Models, products and infrastructure, with source links and review status.
138 entries, including 15 marked for review. Compiled 25 Sept 2026.
Showing 138 of 138 entries
Security and safety toolingsecurity
Proofpoint
Agentic Data and AI Security
Proofpoint's unified system reasoning across AI and data risk in one shared graph, using three autonomous agents to detect, investigate, and remediate threats.
Details and sources
Proofpoint announced the Agentic Data and AI Security System on September 22, 2026, at Proofpoint Protect 2026 in San Diego, describing it as the industry's first unified system reasoning across both AI risk and data risk in one shared graph rather than treating them as separate problems. It runs on three autonomous agents: Zero-Touch Detection, Instant Investigation, and Protection Optimization. Companion capabilities, Semantic Business Policies and Agentic Insights, are expected to be available by year-end 2026.
Key features
Single shared graph reasoning across AI risk and data risk together
Zero-Touch Detection autonomous agent
Instant Investigation autonomous agent
Protection Optimization autonomous agent
Companion Semantic Business Policies and Agentic Insights capabilities (year-end 2026 availability)
Use cases
Detecting and linking risky AI agent behavior to the data it touches
Automated investigation of AI/data security incidents
Continuous optimization of data protection policies in agentic environments
Pricing: unknown; enterprise security platform, most capabilities expected by year-end 2026
An open-source 'Open Agent Operation System' for running and managing agents, now shipping a native macOS app alongside its engine.
Details and sources
AgentOS (use-agent-os/agent-os), described by its maintainers as 'The Open Agent Operation System,' stores its configuration and data under ~/.agentos/, protected across upgrades. Release 2026.9.24, published September 24, 2026, is the first version to ship a native, signed and notarized AgentOS app for macOS (Apple Silicon and Intel) alongside the core engine, with sixty-five pull requests merged since the previous version. The macOS app starts the local gateway if one isn't already running, or adopts an existing gateway started from a terminal, and never starts a second one.
Key features
Signed and notarized native macOS app (Apple Silicon and Intel) as of release 2026.9.24
Local gateway auto-start or adoption of an already-running gateway (never duplicates it)
Configuration and data stored under ~/.agentos/, preserved across upgrades
65 pull requests merged in the 2026.9.24 release cycle
Use cases
Running and managing local AI agents from a native macOS app instead of the terminal
Operating a persistent local agent gateway across upgrades without reconfiguration
Self-hosted agent operations for teams wanting an open alternative to closed agent platforms
An open-source Python framework for building AI agents and multi-agent teams with knowledge retrieval, MCP support, and multi-provider models.
Details and sources
Agno is an open-source framework for building AI agents and agent teams, with session management, workflow orchestration, Model Context Protocol (MCP) support, and multi-provider model access. It moved to the Apache 2.0 license with its v2.5.0 release (February 2026), which added Team Modes and human-in-the-loop support for teams. Its v2.6.10 release (June 2026) added a knowledge-retrieval pipeline with an MMR reranker, a recency reranker blending search scores with timestamp decay, typed page-tool results with explicit error/completeness metadata, and Y-API support as an OpenAI-compatible model provider.
Key features
Multi-agent team orchestration with configurable Team Modes and human-in-the-loop support
Knowledge-retrieval pipeline with MMR reranker for relevance/diversity balance
Recency reranker blending search relevance with timestamp decay
Model Context Protocol (MCP) support and multi-provider model access
Typed page-tool results with explicit error and completeness metadata
Use cases
Building production multi-agent systems with shared knowledge retrieval
Coordinating human-in-the-loop review over agent teams
RAG-style knowledge workflows needing both relevance and recency ranking
Pricing: Free and open source under the Apache 2.0 license.
An open 9B LoRA adapter that scores structured choices/booleans/rubric levels, matching 90% of a 27B model's accuracy on a held-out benchmark.
Details and sources
Bespoke Labs released Bespoke-Nimble-9B on September 18, 2026, a Qwen3.5-9B LoRA adapter trained via contrastive data curation on 2,676 samples in two days, plus the open dataset and training recipe. Rather than generating open-ended text, it takes text and a flat schema (choice, true/false, or rubric score) and returns the picked answer with a calibrated probability for each allowed option, without writing reasoning. On a 324-example internal holdout it matched 90.12% of reference labels, versus 93.21% for the 27B Jev model and 66.36% for the unmodified 9B base model, without distilling from Jev.
Key features
Qwen3.5-9B LoRA adapter (~165 MiB), trained without distillation from a larger model
Contrastive data curation: modifying specific facts to flip correct answers, teaching the model to isolate evidence
Softmax scoring at T=1.0 with probability-weighted expected scores for rubric fields
Supports up to 255 choices per field, 8,192-token context window
90.12% accuracy on a 324-example holdout vs. 93.21% for the 27B Jev model
Open dataset, LoRA weights, and training recipe released together
Use cases
Structured classification, boolean judgments, and rubric scoring at lower cost than a larger model
Consistency evaluation and decision-making tasks needing calibrated per-option probabilities
Cheap local replacement for larger judge/classifier models in constrained-answer settings
Pricing: Free and open source (Apache 2.0); requires a separately hosted Qwen3.5-9B base model and a CUDA GPU with BF16 support.
Google's AI-generated summaries and direct answers for natural-language questions typed into Gmail's search bar, now available globally.
Details and sources
Google announced global availability of Gmail Search AI Overviews on September 15, 2026, letting users ask natural-language questions in Gmail's search bar and get concise summaries and answers without opening individual emails. The feature expanded from a US-only launch to global availability in English, rolling out via Rapid Release starting September 3, 2026 (up to 15 days) and Scheduled Release starting September 21, 2026 (1-3 days for full rollout). It requires Gemini for Workspace and Workspace Intelligence to be enabled by admins, with smart features turned on for users, and is available across Business, Enterprise, Education, Frontline Plus, and consumer Google AI Plus/Pro/Ultra plans (excluding the EEA, UK, Switzerland, and Japan for consumer plans).
Key features
Natural-language question answering directly in Gmail's search bar
AI-generated summaries and direct answers across matching emails
Global rollout in English (previously US-only)
Available across Business, Enterprise, Education, Frontline Plus, and consumer Google AI Plus/Pro/Ultra plans
Use cases
Answering questions like "when is my flight" or "what did the vendor quote us" without opening threads
Summarizing information scattered across multiple emails
Faster email search for Workspace admins who enable Gemini and Workspace Intelligence
Pricing: Included with eligible Business/Enterprise/Education/Frontline Plus Workspace plans and consumer Google AI Plus, Pro, and Ultra plans (excluding EEA, UK, Switzerland, Japan for consumer plans); requires admin-enabled Gemini for Workspace and Workspace Intelligence.
Nous Research's open-source, self-improving AI agent with a closed learning loop and access via CLI, desktop app, and chat platforms.
Details and sources
Hermes Agent is Nous Research's open-source, self-improving agent platform, built around a closed learning loop that creates skills from experience and improves them during use, with persistent memory and full-text search across past conversations. It ships as a unified gateway across CLI, Telegram, Discord, Slack, WhatsApp, Signal, and Email, plus a native desktop app for macOS, Windows, and Linux (public preview from June 2, 2026). The August 3, 2026 'Herald Release' added conversational voice, on-device wake words, grounded research citations, signed outbound webhooks, and A2A support, and Nous later shipped 'Bot Mode,' replacing the single-agent session list with a roster of named bots. It is MIT-licensed, supports Nous Portal, OpenRouter, OpenAI, and custom model endpoints, and had crossed 214,000 GitHub stars by September 2026, which Nous calls the fastest-growing open-source agent framework of the year.
Key features
Closed learning loop: creates and improves skills from experience, with persistent cross-session memory
Unified gateway across CLI, Telegram, Discord, Slack, WhatsApp, Signal, and Email
Native desktop app (macOS/Windows/Linux) in public preview since June 2, 2026
Herald Release (Aug 3, 2026): conversational voice, on-device wake words, grounded citations, signed webhooks, A2A support
Bot Mode: roster of named bots replacing the single-agent session list
Built-in cron scheduler for unattended tasks; flexible model backend (Nous Portal, OpenRouter, OpenAI, custom endpoints)
Use cases
Personal or team AI assistant reachable from existing chat apps
Self-improving automation that accumulates reusable skills over time
Scheduled/unattended agent tasks via built-in cron
Multi-bot workflows via Bot Mode
Pricing: Free and open source (MIT license); model usage costs depend on the connected backend (Nous Portal, OpenRouter, OpenAI, or custom endpoint).
An experimental ~280M-parameter draft model that speeds up LFM2.5-VL-3B vision-language inference up to 3.13x via speculative decoding.
Details and sources
Liquid AI released LFM2.5-VL-DSpark on September 24, 2026, an experimental draft model for its LFM2.5-VL-3B vision-language model implementing speculative decoding: a small draft model proposes candidate tokens that the larger target model verifies. The ~280M-parameter draft uses a 4-layer attention-only decoder with a block size of 8-9, delivering exact speculative decoding (output identical to running the 3B target alone) with decode speedups up to 3.13x on-device and 2.66x on an H100, and end-to-end speedups of 2.62x on-device and 2.27x on GPU. It ships with day-one support for llama.cpp, MLX-VLM, and SGLang, and was evaluated on six vision tasks (VQA, text VQA, captioning, chart QA, reasoning, multi-turn conversation) under the MMSpec benchmark standard.
Key features
~280M-parameter draft model (4-layer attention-only decoder) pairing with LFM2.5-VL-3B
Exact speculative decoding: output matches running the 3B target model alone
Decode speedups up to 3.13x on-device, 2.66x on H100
End-to-end speedups of 2.62x on-device, 2.27x on GPU
Day-one support for llama.cpp, MLX-VLM, and SGLang
Open weights in Safetensors and GGUF formats, unrestricted for download, fine-tuning, and deployment
Use cases
Faster on-device vision-language inference without changing model outputs
Lower-latency VLM deployment on llama.cpp, MLX-VLM, or SGLang
Cost reduction for GPU-served vision-language inference via speculative decoding
Pricing: Free; open weights, unrestricted for download, fine-tuning, and deployment.
Infrastructure and inference/servinginfrastructure
llama.cpp
llama.cpp v0.5.0
The v0.5.0 release of the widely used C/C++ LLM inference engine, adding new model support and backend performance work.
This entry is marked for review.
Details and sources
llama.cpp shipped v0.5.0, following v0.4.0 (September 4, 2026) and v0.4.1 (September 14, 2026) under the project's recently adopted semantic-versioning scheme (the project moved from continuous build-tagged releases to numbered semver releases starting August 25, 2026). The release focuses on backend performance and correctness, broader model coverage, and more robust server/router operation: CUDA conv2d acceleration via implicit GEMM, Metal MoE and SSM_CONV fusion, multi-address HTTP binding for servers, new model support (HRM-Text/DFM Mimir 1B, MiMo-V2.6 conversion, HunyuanOCR with DFlash, extended Nemotron and Qwen4 handling), an updated ggml v0.25.0 with expanded hyper-connection and flash-attention support, and function-call outputs that now support image input.
Key features
CUDA conv2d acceleration via implicit GEMM; Metal MoE and SSM_CONV fusion
Multi-address HTTP binding for servers (comma-separated TCP addresses and UNIX sockets)
New model support: HRM-Text (DFM Mimir 1B), MiMo-V2.6 conversion, HunyuanOCR with DFlash, extended Nemotron/Qwen4 handling
Updated ggml to v0.25.0 with expanded hyper-connection and flash-attention support
Function-call outputs now support image input
LoRA loading from open file pointers
Use cases
Local and server-side LLM/VLM inference in C/C++ with minimal dependencies
Cross-hardware deployment (CPU, Apple Silicon, CUDA, HIP, MUSA, Vulkan, SYCL)
Running newly supported model families (MiMo-V2.6, HunyuanOCR, Nemotron, Qwen4) locally
Pricing: Free and open source under the MIT license.
OpenHands' open-source autonomous coding-agent platform, with v1.23.0 adding a universal macOS DMG and new UI themes.
Details and sources
OpenHands is an open-source AI agent platform for autonomous software development, letting agents interact with code repositories, execute tasks, and work through development workflows locally, in Docker, or in the cloud, with support for multiple LLM providers and version control systems. Version 1.23.0, released September 23, 2026, adds a universal macOS DMG with per-architecture bundled runtimes, Light+ and Solarized Light UI themes, Agent Canvas telemetry improvements, and bug fixes for Canvas image packages, chat rendering, and automation scripts.
Key features
Autonomous coding agent that edits repos, runs commands, and executes multi-step dev tasks
Universal macOS DMG with per-architecture bundled runtimes (v1.23.0)
Light+ and Solarized Light themes added (v1.23.0)
Agent Canvas telemetry improvements
Local, Docker, and cloud deployment options with multi-LLM-provider support
Use cases
Autonomous end-to-end coding tasks (bug fixes, feature implementation) against a real repository
Local or cloud-hosted coding-agent workflows integrated with existing VCS
Desktop macOS coding-agent usage via the new native DMG installer
Pricing: Free and open source; hosted/cloud usage costs depend on the connected LLM provider.
Amazon's AI agent for marketplace sellers, now with persistent memory across sessions and always-on autonomous restocking/pricing workflows.
Details and sources
Amazon announced the next evolution of Seller Assistant on September 23, 2026, at Amazon Accelerate 2026, adding persistent memory of each seller's pricing patterns, inventory cycles, and growth goals that follows sellers across Seller Central, Amazon Quick, and Claude instead of resetting between sessions. New always-on workflows continuously monitor conditions and execute routine tasks such as restocking and pricing adjustments in the background, even when the seller isn't logged in, with full audit trails and seller-defined guardrails. A new Selling Partner plugin, launching with Amazon Quick and in beta with Anthropic's Claude, connects Seller Assistant to those external AI tools in roughly 60 seconds with no coding required.
Key features
Persistent memory of pricing patterns, inventory cycles, and growth goals across Seller Central, Amazon Quick, and Claude
Always-on background workflows for restocking and pricing adjustments with full audit trails
Seller-defined guardrails on autonomous actions
Selling Partner plugin for Amazon Quick (launch) and Claude (beta), connecting in about 60 seconds with no coding
Example workflow: alerting when a top product's rating drops below 4 stars, or auto-adjusting price when a competitive opening appears
Use cases
24/7 automated restocking and pricing adjustments for marketplace sellers
Cross-session seller assistant memory instead of resetting context each visit
Accessing Amazon seller data and Seller Assistant intelligence from inside Claude or Amazon Quick
Unsloth's open-source fine-tuning toolkit, now spanning diffusion, TTS, and embedding models with a native desktop app and self-hosted web UI.
Details and sources
Unsloth is an open-source library for fast, memory-efficient local fine-tuning of language models, which has expanded through 2026 to also cover diffusion, text-to-speech, and embedding models, alongside a native desktop app and a self-hosted web UI. Its September 2026 releases (v2026.9.4 on September 9 and v2026.9.23 on September 23) added support for exporting NVFP4, FP8, and imatrix GGUFs after training, the ability to act as a llama-swap API system, and Japanese and Brazilian Portuguese localization.
Key features
Fast, memory-efficient fine-tuning for LLMs, diffusion, TTS, and embedding models
Native desktop app and self-hosted web UI
Export to NVFP4, FP8, and imatrix GGUF formats after training
Can act as a llama-swap API system
Japanese and Brazilian Portuguese localization added in September 2026
Use cases
Local, low-memory fine-tuning of open LLMs and diffusion/TTS/embedding models
Exporting fine-tuned models directly to quantized GGUF formats for local inference
Self-hosted fine-tuning workflows via the desktop app or web UI
D-Robotics AI Lab's open robot simulator that autoregressively generates video observations from joint-trajectory input at 24 FPS.
Details and sources
D-Robotics AI Lab published Uranus (arXiv:2609.24815, v3 dated September 23, 2026), a data-driven robot simulator built around a joint-trajectory-conditioned autoregressive diffusion model. It takes future joint-position trajectories as streaming input and autoregressively generates one latent frame per step (corresponding to four RGB frames) without a fixed horizon, reaching 24 FPS after inference optimization, and provides a unified interface for synchronized multi-view generation across different robot embodiments and camera configurations. The authors released code, model weights, and inference tools openly alongside quantitative and qualitative evaluation on both in-distribution and out-of-distribution data.
Key features
Joint-trajectory-conditioned autoregressive diffusion model for video generation
Streaming rollout: one latent frame (four RGB frames) generated per step, no fixed horizon
24 FPS after inference optimization
Unified interface for synchronized multi-view generation across robot embodiments and camera setups
Open code, model weights, and inference tools
Use cases
Generating realistic visual rollouts for robot policy training and evaluation
Testing robot control policies against out-of-distribution simulated scenarios
Multi-camera, multi-embodiment simulation without per-setup custom tooling
Pricing: Free and open source; code, weights, and inference tools released publicly.
Infrastructure and inference/servinginfrastructure
OpenRouter
Batch API
OpenRouter's asynchronous API for bundling requests across 70+ models at roughly half the standard per-token price.
Details and sources
OpenRouter launched its Batch API on September 22-23, 2026, letting developers submit non-urgent workloads across more than 70 models at roughly half the standard per-token price in exchange for results within a 24-hour window. Across 230,000+ batches completed during a two-week beta, the median batch finished in 7 minutes and 90% finished within an hour. It supports chat completions, responses, messages, and embeddings endpoints, with 30-day data retention and per-request results so individual failures don't affect the rest of a batch.
Key features
70+ supported models with roughly 50% (sometimes more) off standard per-token pricing
24-hour completion window; beta data shows median 7-minute turnaround, 90% within an hour
Covers chat completions, responses, messages, and embeddings endpoints
Per-request results so individual failures don't fail the whole batch
Compatible with Bring Your Own Key (BYOK) and provider allowlist/data policies
Use cases
Labeling large datasets
Backfilling embeddings
Scoring evaluation sets
Summarizing ticket or document backlogs
Running identical prompts across thousands of rows overnight
Pricing: Roughly 50% off standard per-token pricing for most of the 70+ supported models (web search tool calls remain at standard rates); exact discount varies by model.
Anthropic's first 5.5-family model, matching Claude Fable 5.1 on most work at 40% lower cost.
Details and sources
Anthropic released Claude Opus 5.5 on September 22, 2026, the first model in its new 5.5 family and the successor to Claude Opus 5. Anthropic says it performs at the level of Claude Fable 5.1 on most work while costing 40% less to run, with output speeds up over 30% versus Opus 5. On benchmarks Anthropic shared, Opus 5.5 surpassed Fable 5.1 on agentic coding, knowledge work, computer use, visual chart recognition, and multidisciplinary reasoning. It is available on AWS, Google Cloud, and Microsoft Azure, and Anthropic has signaled Sonnet 5.5 and Haiku 5.5 will follow within weeks.
Key features
Matches Claude Fable 5.1 on most work at roughly 40% lower cost than Opus 5
Output speed increased over 30% versus Opus 5
Beats Fable 5.1 on Anthropic's reported agentic coding, knowledge work, computer use, and visual chart recognition benchmarks
Cache reads priced at $0.20/MTok, a 60% reduction from Opus 5
Use cases
Cost-sensitive agentic coding and computer-use workloads
Knowledge work and multidisciplinary reasoning tasks previously requiring Fable-tier pricing
Cloud-hosted deployment via AWS, Google Cloud, or Microsoft Azure
Pricing: $4/MTok input, $20/MTok output (20% below Opus 5), $0.20/MTok cached reads (60% below Opus 5).
Base Compute's open-source leaderboard and CLI for benchmarking local inference across models, quantizations, chips, and runtimes.
Details and sources
Base Compute released Compute:Arena on September 22, 2026, a public leaderboard and CLI for measuring local inference performance across the combination of model, quantization format, hardware chip, and runtime. It measures prefill throughput (128-16,384 tokens) and decode throughput (128 tokens), records model file hashes, quantization formats, and chip specs, and generates signed local reports capturing telemetry like temperature, power state, and memory pressure. It supports two runtimes, BaseRT for .base bundles and llama.cpp for .gguf files, verifies benchmarks offline before submission, and signs reports with Ed25519 to ensure integrity.
Key features
Benchmarks the combined model/quantization/chip/runtime 'four-dimensional problem' rather than one axis at a time
Measures prefill (128-16,384 tokens) and decode (128 tokens) throughput
Captures telemetry: temperature, power state, memory pressure
Ed25519-signed local reports for verifiable, tamper-evident results
Supports BaseRT (.base) and llama.cpp (.gguf) runtimes
Use cases
Comparing local inference performance across hardware before deployment
Verifying quantization/runtime tradeoffs with signed, reproducible reports
Public leaderboard benchmarking for local/edge inference setups
Pricing: Free and open source under the Apache 2.0 license; CLI available at computearena.ai.
Evaluation, observability, and MLOps toolingframework
EvalEval Coalition
Evaluation Cards
A shared reporting schema and open platform, built with the UK AI Security Institute, for making AI benchmark results comparable across setups.
Details and sources
The EvalEval Coalition, working with the UK AI Security Institute (AISI), launched Evaluation Cards on September 22, 2026, an open platform and shared reporting schema that consolidates benchmark metadata, evaluation-run data, and model information into unified, comparable records. It uses the 'Every Eval Ever' (EEE) schema for structured documentation, provides transcript-level transparency for reproducibility, and surfaces contextual information about how setup choices affect reported performance. Model developers can submit verified results, evaluation developers can report benchmarks using the EEE schema, and researchers can browse cards by benchmark or model at evalcards.evalevalai.com.
Key features
Shared 'Every Eval Ever' (EEE) reporting schema across benchmarks and models
Transcript-level transparency for reproducibility and analysis
Contextual notes on how evaluation setup choices affect reported scores
Verified result submission for model developers
Browsable by benchmark or by model
Use cases
Comparing benchmark results across models and evaluation setups on equal footing
Auditing evaluation methodology via transcript-level detail
Standardized benchmark reporting for evaluation-tooling developers
Pricing: Free to use; available at evalcards.evalevalai.com.
A 7B-parameter open-weight world action model that turns camera frames and text instructions into the next 2 seconds of robot actions.
Details and sources
Black Forest Labs released FLUX 3 Action on September 23, 2026, a 7-billion-parameter open-weight world action model distinct from its FLUX 3 image generator. It takes multiple camera angles, state vectors, and text captions as input and jointly predicts future video frames and robot movements using a diffusion transformer built on a frozen video VAE and Qwen3-VL encoders, outputting 32 actions with optional decoded frames. It is fine-tunable for robotics and non-robotics tasks like games and drones, integrates with LeRobot for robot-arm deployment, and scored 42.92% success on RoboLab-120, ahead of Cosmos 3 Nano's 36.8% despite Cosmos 3 Nano having more than double the parameters (16B).
Key features
7B-parameter diffusion transformer combining a frozen video VAE with Qwen3-VL encoders
Predicts the next 2 seconds of actions (32 actions) from camera frames, state vectors, and text captions
Optional decoded future video frames alongside predicted actions
Fine-tunable for robotics, games, and drone control
Native LeRobot integration for robot-arm deployment
Recovers from its own mistakes and adapts to unseen objects/camera positions
Use cases
Robot-arm manipulation and grasping via LeRobot deployment
Fine-tuning for game or drone control from visual input
Research into joint video-and-action world models
Pricing: Free; open weights released under the FLUX Kommunity License v1.0.
Google's most expressive text-to-speech model, generating custom voices from natural-language prompts across 100+ languages.
Details and sources
Google released Gemini 3.8 Flash TTS on September 23, 2026, as its most expressive audio generation model to date. It supports generative voice design from natural-language prompts, voice replication from 30-second samples with consent verification and SynthID watermarking, line-by-line performance direction, hours-long consistent long-form narration, dual-speaker scene staging, and non-verbal vocal bursts like laughs and sighs. It ranked #1 overall on Hume AI's Voice Design Benchmark (71.4) and topped Voice Arena leaderboards for Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish, and Hindi.
Key features
Generative voice design from natural-language prompts across 100+ languages and dialects
Voice replication from 30-second audio samples with consent verification and SynthID watermarking
Line-by-line performance direction (pacing, emotion, delivery via stage directions)
Long-form generation maintaining voice/character consistency across hours of audio
Dual-speaker scene staging and realistic non-verbal vocal bursts
Google's lower-cost sibling to Gemini 3.8 Flash TTS, bringing expressive voice generation to Google Vids.
Details and sources
Google released Gemini 3.8 Flash-Lite TTS alongside Flash TTS on September 23, 2026, as a lighter-weight audio generation model sharing the same generative voice design, voice replication, and performance-direction capabilities. It ranked #2 on Hume AI's Overall Quality Index behind Flash TTS. It ships to general users inside Google Vids, while developers get it via the Gemini API and Google AI Studio.
Key features
Generative voice design from natural-language prompts across 100+ languages and dialects
Voice replication from 30-second audio samples with consent verification and SynthID watermarking
Line-by-line performance direction and non-verbal vocal bursts
Access to the same 2,000+ production-ready voice library as Flash TTS
Ranked #2 on Hume AI's Overall Quality Index
Use cases
Lower-cost voiceover generation inside Google Vids
Custom voice creation for lighter-weight production workflows
Developer integrations needing expressive TTS at a lower price point than Flash TTS
Pricing: unknown; available today via the Gemini API and Google AI Studio for developers, and inside Google Vids for general users
GitHub's CLI extension that compiles Markdown-defined AI automation into standard, sandboxed GitHub Actions workflows.
Details and sources
GitHub Agentic Workflows (gh-aw), a collaboration among GitHub Next, Microsoft Research, and Azure Core Upstream, lets developers define AI-powered repository automation in Markdown with YAML frontmatter, which the CLI compiles into standard GitHub Actions workflows. It entered technical preview in February 2026 and moved to public preview on June 11, 2026. Agent jobs are read-only and sandboxed by default, with any writes handled through separate jobs with scoped permissions, and it supports GitHub Copilot, Claude Code, OpenAI Codex, Google Gemini, and Pi as AI engines for tasks like issue triage, PR review, CI failure investigation, documentation maintenance, and dependency analysis. It is MIT-licensed and installed via `gh extension install github/gh-aw`; a security vulnerability affecting versions 0.83.3 to 0.85.3 was patched in 0.85.4.
Key features
Markdown + YAML frontmatter workflow definitions compiled into standard GitHub Actions
Read-only, sandboxed agent jobs by default; writes go through separately scoped jobs
Multi-engine support: GitHub Copilot, Claude Code, OpenAI Codex, Google Gemini, Pi
Complements rather than replaces existing CI/CD workflows
MIT-licensed CLI extension
Use cases
Automated issue triage and pull-request review
CI failure investigation requiring reasoning, not just deterministic checks
Documentation maintenance and dependency analysis automation
Repository reporting driven by natural-language workflow definitions
Pricing: Free and open source (MIT license); billed usage depends on the underlying AI engine and GitHub Actions minutes consumed.
Alphabet robotics unit Intrinsic's open-sourced, ROS-compatible foundation for industrial robot control, perception, and motion planning.
Details and sources
Intrinsic, Alphabet's robotics company, open-sourced Intrinsic Core under the Apache 2.0 license at ROSCon 2026 in Toronto on September 22, 2026. The release bundles Intrinsic Control (a hardware-agnostic real-time control framework), pose estimation built on Nvidia's FoundationPose, motion and grasp planning, simulation and calibration services, and Intrinsic-ROS drivers for third-party sensor and hardware integration, all runnable in a pre-configured environment on local hardware. Alongside Intrinsic Core, Intrinsic released the Open Machine Tending Solution, an open reference design for AI-powered CNC machine tending supporting Universal Robots and FANUC equipment.
Key features
Apache 2.0-licensed, ROS-compatible robotics stack, code on GitHub
Hardware-agnostic real-time control framework (Intrinsic Control)
Pose estimation built on Nvidia FoundationPose
Motion planning, grasp planning, simulation, and calibration services included
Companion Open Machine Tending reference design for Universal Robots and FANUC hardware
Use cases
Building custom industrial robotic applications on ROS-compatible hardware
Robot pose estimation and grasp/motion planning without proprietary lock-in
AI-powered CNC machine tending using Universal Robots or FANUC arms
Pricing: Free and open source under the Apache 2.0 license.
JetBrains' cross-vendor system for coordinating agentic software development inside and beyond its IDEs.
Details and sources
JetBrains introduced JetBrains Air on September 22, 2026, an open, coherent system of products for coordinating agentic software development across developers, teams, and organizations, inside and beyond JetBrains IDEs. It comprises Air in JetBrains IDEs (directing agents and verifying their work with code intelligence), Air Teams (coordinating workflows between developers and autonomous agents), Air Governance (organizational policy, visibility, cost management, and accountability), the Agent Client Protocol (ACP) for standardizing connections between IDEs and agents from multiple vendors, and Junie, JetBrains' own coding agent supported across all Air surfaces. JetBrains says Air will roll out through a series of releases, with some components available now and others in preview.
Key features
Air in JetBrains IDEs for directing and verifying agent work with code intelligence
Air Teams for coordinating developer/agent workflows
Air Governance for org-wide policy, cost, and accountability controls
Agent Client Protocol (ACP) standardizing IDE-to-agent connections across vendors
Junie, JetBrains' own coding agent, available across all Air surfaces
Use cases
Directing and reviewing AI coding agents from within JetBrains IDEs
Coordinating mixed human/agent development teams
Enforcing organizational governance over agentic coding tools
Connecting third-party agents to IDEs via a standard protocol
MiniMax's open-source terminal coding agent, usable with MiniMax models or any OpenAI/Anthropic-compatible provider.
Details and sources
MiniMax open-sourced MiniMax Code on September 18, 2026, a terminal coding agent that understands projects, edits files, runs tests, and executes shell commands from the command line. It offers both an interactive terminal UI and a headless CLI mode for scripts and CI, works with MiniMax accounts or custom OpenAI/Anthropic-compatible API providers, and integrates with editors via the Agent Client Protocol (ACP). The repository ships as a v0.4.12 source preview under an MIT license (with dependency licenses preserved) and had reached 1.8k GitHub stars and 206 forks shortly after release.
Key features
Interactive terminal UI and headless CLI modes
Works with MiniMax accounts or bring-your-own OpenAI/Anthropic-compatible API keys
File editing, diff inspection, shell command execution, and test running
Session resumption and task planning
Agent Client Protocol (ACP) integration for editors
Built-in search, plugins, and multimodal tools
Use cases
Terminal-based project understanding and code editing
CI/scripted headless coding-agent runs
Bring-your-own-model coding workflows via OpenAI/Anthropic-compatible APIs
Pricing: Free and open source (MIT license) on GitHub; usage requires either a MiniMax account with credits or a bring-your-own API key.
A 100M-parameter open-weight speaker diarization model handling up to 8 speakers for both live and recorded audio.
Details and sources
NVIDIA released Nemotron 3 Diarization on September 23, 2026, a 100-million-parameter open-weight model that assigns anonymous speaker labels to overlapping speech, expanding support from 4 to 8 speakers versus its predecessor. It offers four configurable latency modes (30.4s, 1.04s, 0.64s, 0.32s) and works as one model for both offline and streaming conversations. It ranked #1 on VoiceArena's Diarization-Bench with a 14.72% Diarization Error Rate, a 40% average relative DER reduction versus NVIDIA's previous baseline at 1.04-second latency, and a claimed 15,113x real-time factor speedup at batch size 32.
Key features
Supports up to 8 speakers, up from 4 in the prior generation
Four configurable latency modes from 0.32s to 30.4s
Single model for both offline and streaming diarization
Handles overlapping speech natively
Chunked processing for flexible recording lengths
Use cases
Meeting and call transcription with speaker labeling
Real-time streaming diarization for live conversations
Pricing: Free; open weights on Hugging Face under the OpenMDW License Agreement v1.1. Hosted integration available via Baseten, DigitalOcean, and Argmax Pro SDK 3.
Rabbit's agentic operating system that orchestrates AI models, web services, and up to 5 devices from a single chat interface.
Details and sources
Rabbit launched OS3 on September 22, 2026, an agentic operating system controlled through a single continuous chat interface with memory across sessions. Users state a desired outcome and OS3 determines how to accomplish it across connected AI models, web services, and devices, automatically choosing the optimal device for each task. It supports Bring Your Own Key model swapping, universal skill integration by pasting URLs, direct computer control and autonomous coding via rabbit's DLAM technology, and local file processing without storing or copying data. It runs across Windows, Mac, and Linux PCs, cloud VMs, dedicated AI machines, the r1 device, and an upcoming cyberdeck, with one account supporting up to five connected devices.
Key features
Single continuous chat interface with cross-session memory
Multi-device orchestration that auto-selects the optimal device per task
Bring Your Own Key support to swap between AI models
Universal skill integration via pasted URLs
Direct computer control and autonomous coding via DLAM technology
Local file processing without storing or copying data
Use cases
Cross-device task automation from one chat interface
Autonomous coding and computer control
Personal assistant workflows spanning PCs, cloud VMs, and the r1 device
Hugging Face's open-source repository-memory server that lets coding agents query GitHub issue/PR history alongside a working clone.
Details and sources
Hugging Face released relore on September 21, 2026, an Apache 2.0-licensed tool that indexes a repository's GitHub issue, pull-request, and comment history so coding agents can connect past decisions to current code. It keeps a working clone beside the history index and serves both over one HTTP API, distinguishing maintainer decisions from contributor claims and automated comments rather than treating all evidence equally, and excludes machine-generated content from default search results. It ships specialized commands including `inflight` (in-progress fixes), `thread` (discussion access), `why` (line-origin explanations via git blame and reviews), and code-browsing verbs like `defs` and `copies`. Queries run locally; only ingestion contacts GitHub. It already runs internally across several Hugging Face projects, including Transformers.
Key features
Indexes GitHub issues, PRs, and comments alongside a working repo clone
Single HTTP API serving both code and history
Distinguishes maintainer decisions from contributor claims and bot comments
StepFun's 600B-parameter (27B active) sparse MoE flagship for long-horizon agentic and coding work, priced at $1/M input tokens.
Details and sources
StepFun launched Step 5 Preview on September 20, 2026, a 600-billion-parameter sparse mixture-of-experts model with 27B parameters active per token, a 1M-token context window, and native text, image, and video input, aimed at software engineering, professional knowledge work, and finance. It scored 44 on the Artificial Analysis Intelligence Index against a similar-price-tier median of 24. In a 24-hour agent experiment, it tuned an H100 kernel to 508 TFLOPS versus 493 for Claude Opus 5, and separately raised Qwen3-30B-A3B's AIME24 score from 53.3% to 60% through automated post-training. StepFun opened paid API access the same day and has a countdown for open weights on October 15, 2026.
Key features
600B total parameters, 27B active per token (sparse MoE)
1M-token context window with native text, image, and video input
Artificial Analysis Intelligence Index score of 44 vs. a similar-price-tier median of 24
Demonstrated automated kernel tuning (508 TFLOPS on H100) and automated post-training gains in agent experiments
Open weights planned for October 15, 2026
Use cases
Long-horizon agentic software engineering tasks
Professional knowledge work and finance workflows
Automated ML engineering tasks like kernel tuning and post-training
Pricing: $1.00/million input tokens (cache miss), $0.05/million (cache hit), $2.70/million output tokens (reasoning tokens billed as output) via StepFun's API.
A release-candidate rewrite of Hugging Face's Tokenizers library that is 3-30x faster while producing identical token IDs to v0.23.
Details and sources
Hugging Face published a release candidate of Tokenizers v1.0.0 on September 21, 2026, producing the same token IDs as v0.23 while running 3 to 30 times faster single-threaded on an Apple M4 Max. The rewrite introduces 'Bitcannon,' which replaces regex splitting with SIMD bitstream operations for GPT-2, cl100k, o200k, Tekken, and DeepSeek tokenizer families; a thread-local word cache that avoids reprocessing repeated words; a workspace split so applications link only the components they need; native parallelism without single-lock queuing; and reusable scratch buffers that cut repeated allocations. The full 1.0.0 release is planned to add more model families, simplified Python bindings, and inference-only C/C++ bindings.
Key features
3-30x single-threaded speedup vs. v0.23 while preserving identical token IDs
Bitcannon: SIMD bitstream splitting replacing regex for GPT-2, cl100k, o200k, Tekken, and DeepSeek tokenizers
Thread-local word cache avoiding reprocessing of repeated words
Workspace split so apps link only needed components
Native parallelism without single-lock thread queuing
Use cases
High-throughput tokenization for LLM training and inference pipelines
Rust applications wanting a lighter-weight tokenizer dependency
Migration path to a faster tokenizer with no output drift from v0.23
Pricing: Free and open source; install the release candidate via `cargo add tokenizers --pre`.
Alibaba's AI training and inference chip with 216GB of memory, built to support 500,000-chip superclusters for future Qwen models.
Details and sources
Alibaba's chip subsidiary T-Head unveiled the Zhenwu V900 AI accelerator at the Apsara Conference 2026 in Hangzhou on September 22, 2026. Alibaba CEO Eddie Wu called it the most powerful AI chip in China, claiming three times the performance of the prior-generation M890. It has 216GB of GPU memory, 1,200GB/s of inter-chip bandwidth, and native FP8/FP4 low-precision compute support, and is designed to scale into clusters of up to 500,000 accelerator cards to train future Qwen models in the 5-10 trillion parameter range. Mass production is targeted for Q1 2027, alongside a plan to expand Alibaba Cloud's global data center capacity past 20 gigawatts by 2032.
Key features
216GB GPU memory, 1,200GB/s inter-chip bandwidth
Native FP8 and FP4 low-precision compute support
Claimed 3x performance of Alibaba's prior-generation M890 chip
Scales to superclusters of up to 500,000 accelerator cards
Built to train Qwen models targeting 5-10 trillion parameters
Use cases
Large-scale AI model training within Alibaba Cloud infrastructure
High-bandwidth inference serving for frontier-scale models
China-domestic alternative to export-restricted Nvidia accelerators
Google's Kotlin/Android agent framework reaches 1.0 with feature parity to ADK Core plus on-device and Android-first extensions.
Details and sources
Published September 9, 2026, ADK for Kotlin 1.0 reaches full feature parity with ADK 1.0 Core (already available for Python, Go, Java, and TypeScript) and adds Android-first extensions: on-device agents via LiteRT-LM and ML Kit (beta), hybrid cloud workflows through Firebase AI Logic, and agent-state persistence via Room and AppSearch. It supports hierarchical multi-agent systems where agents delegate to specialized child agents.
Key features
Full feature parity with ADK 1.0 Core
On-device agents via LiteRT-LM and ML Kit (beta)
Hybrid cloud workflows via Firebase AI Logic
Agent-state persistence via Room and AppSearch
Hierarchical multi-agent delegation to child agents
Use cases
Android apps embedding on-device or hybrid AI agents
Kotlin/JVM backend services building multi-agent systems
Kotlin Multiplatform agent apps needing parity with Python/Java ADK
Mistral's multi-step retrieval layer that lets models search, open, navigate, read, and grep documents iteratively instead of one-shot RAG.
Details and sources
Agentic Search, announced August 20, 2026, is a retrieval layer built on top of existing search indexes that gives models five tools -- search, open, navigate, read, and grep -- to iteratively find, inspect, and verify information across long or complex documents rather than relying on single-shot RAG retrieval. Mistral reports it triples accuracy on FinanceBench (26.7% to 86% with Mistral Medium 3.5) while cutting token usage by 23.9% and p90 latency from 255s to 154s, and drives a 45.6-point accuracy gain on the table-heavy OfficeQA Pro benchmark. It is built into Mistral's Search Toolkit, Libraries, Studio, and Vibe, and is deployable in cloud or on-premises environments without model fine-tuning.
Skynet.io's task-executing AI Agents plus a persistent Database that gives them shared, cross-run memory.
This entry is marked for review.
Details and sources
On August 4, 2026, Skynet.io (self-described as "the industry's most generous AI app", unrelated to the separate crypto/blockchain agent network operating under the same Skynet name) launched two features together: Agents, AI workers that execute tasks on-demand, on a schedule, or via triggers and report back what they changed rather than just answering questions; and Database, persistent structured storage that lets agents retain information between runs, with multiple agents able to read and write the same records, which users can also edit directly or through chat. Both are available now on Skynet's Pro tier and above. Cited use cases include weekly reporting, support triage, pipeline hygiene, and competitive monitoring.
Key features
Agents: task-executing AI workers triggered on-demand, on schedule, or by events
Database: persistent structured storage giving agents memory across runs
Multiple agents can share and edit the same Database records
Records editable directly or via chat
Available on Pro tier and above
Use cases
Weekly business reporting
Customer support triage
Sales pipeline hygiene
Competitive monitoring
Pricing: Available on Pro and above; exact tier pricing not published in the sourced material. Free registration exists for lower-tier chat/mini-app use.
OpenAI's managed API bringing Codex's agent harness, durable sessions, and hosted sandboxes to developers building production agents.
Details and sources
OpenAI introduced the Agents API in public beta on September 10, 2026, exposing the durable session and orchestration layer behind Codex to third-party developers. It manages context, recovery, and multi-step execution for agents that run continuously in production, with automatic context compaction, multi-agent orchestration, programmatic tool calling, and MCP server support. OpenAI hosts and manages the underlying sandbox by default, but developers can instead run agents on their own infrastructure or through supported sandbox providers including Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel.
Key features
Durable, recoverable multi-step agent sessions with automatic context compaction
Multi-agent orchestration and programmatic tool calling
Evaluation, observability, and MLOps toolingproduct
NIST
AI Technology Evaluation (AITE)
NIST's sequestered testbed for blind-data evaluation of AI models, launching with vision-language tasks in quantum science, genomics, and public safety.
Details and sources
NIST's Technology Test and Evaluation Division announced the Artificial Intelligence Technology Evaluation (AITE) program in July 2026, providing researchers a sequestered testbed environment to evaluate AI model performance on blind data across diverse datasets, modalities, and domains, mitigating train/test contamination risk. NIST provides common data, metrics, and scoring so developers can compare model performance under a consistent, objective standard. AITE launches with three initial tasks: large vision-language model (VLM) image analysis in quantum science, genomics, and public safety, with plans to add more tasks under themes such as Quantum, Video, and NLP over time.
Key features
Sequestered testbed environment for blind-data model evaluation
Mitigates train/test data contamination risk
Common data, metrics, and scoring across submissions
Launch tasks: VLM image analysis for quantum science, genomics, and public safety
Planned expansion into additional themes (Quantum, Video, NLP)
Use cases
Objective, contamination-resistant benchmarking of vision-language models
Comparing model performance on specialized scientific/public-safety imagery tasks
Community AI evaluation research
Pricing: Not applicable (free government-run evaluation program for volunteer participants).
Salesforce's interface layer that brings its CRM data, workflows, and governance into Claude, Slack, and Agentforce.
Details and sources
Salesforce unveiled AIforce at Dreamforce on September 15, 2026, positioning it as an interface layer that carries Salesforce's enterprise data, workflows, business logic, permissions, and governance into whatever surface people or agents already work in, rather than requiring them to open Salesforce itself. The initial set of surfaces includes Claudeforce (Claude), Slackforce (Slack), and Agentforce Coworker, all built on top of Salesforce's existing Data 360, Customer 360, and Agentforce architecture. CEO Marc Benioff framed it as an "interface revolution" comparable to the shift from DOS to GUIs.
Key features
Carries Salesforce data, workflows, permissions, and governance into external interfaces
Amazon's generative-AI Alexa+ assistant now supports MCP for third-party integrations, plus web access and Wallet-based voice purchasing.
Details and sources
In July 2026, Amazon opened new ways to build for Alexa+: an AI-powered smart-home developer toolkit, Model Context Protocol (MCP) support for service integrations, and Amazon Wallet integration for voice purchases -- all in preview. Alexa+ also expanded to web access via Alexa.com and got a redesigned mobile app, after nine months in early access with reported gains in engagement and purchasing activity.
Key features
MCP support for third-party service integrations (preview)
AI-powered smart-home developer toolkit
Amazon Wallet integration for voice-based purchases
Web access via Alexa.com
Redesigned mobile app
Use cases
Voice-driven smart home control across third-party ecosystems (Philips Hue, Roborock)
Voice commerce and reservations (OpenTable, Vagaro)
Browser-based assistant access
Pricing: unknown -- consumer or developer pricing for the new toolkit not disclosed.
Google's Gemini-powered evolutionary code-optimization agent, made generally available on Google Cloud on July 10, 2026.
Details and sources
AlphaEvolve, Google DeepMind's Gemini-powered evolutionary code-optimization agent first introduced in mid-2025, reached general availability on Google Cloud on July 10, 2026, opening access beyond its prior private-preview/early-access program. Users submit a baseline algorithm, problem context, and a scoring function; AlphaEvolve searches large algorithmic spaces and returns human-readable, optimized code through a define-measure-optimize-apply workflow. At GA, Google cited results from 16 early-access organizations, including an 80% planning-model improvement for BASF, a 22% accuracy gain and 90% runtime reduction for Kinaxis, and a 4x speed increase for Schrodinger.
Key features
General availability on Google Cloud as of July 10, 2026 (previously private preview/early access only)
Define-measure-optimize-apply workflow for evolving algorithms against a client-side scoring function
Arize AX made sessions a first-class object and expanded Agent-as-a-Judge evaluation to every pricing tier.
Details and sources
Arize's August 6-September 18, 2026 AX release made conversation sessions a first-class, actionable unit (annotate, queue for review, evaluate at session level) and expanded Agent-as-a-Judge -- which inspects a production trace, finds relevant spans, and labels issues -- from Enterprise-only to every plan. The release also added vision evaluations (LLM judges can reference image columns across traces, datasets, and Playground), a new client.traces.list() Python SDK method, faster dataset population from traces, and support for new models including Claude Fable 5.1, OpenAI's GPT-6/gpt-6-astra, and Gemini 3.7 Flash.
Key features
Sessions as first-class objects (annotate, queue, evaluate at session level)
Agent-as-a-Judge on every plan (previously Enterprise-only)
Vision evaluations across traces, datasets, and Playground
New client.traces.list() SDK method
New model support: Claude Fable 5.1, GPT-6/gpt-6-astra, Gemini 3.7 Flash
Use cases
Evaluating multi-turn agent conversations as a whole
Automated agent trace grading without human review
Multimodal (image/audio) agent evaluation
Pricing: unknown -- Agent-as-a-Judge availability expanded to all tiers, but tier pricing not disclosed.
Google open-sourced AX, an Apache-2.0 Kubernetes-style declarative runtime for orchestrating autonomous AI agents at scale.
Details and sources
Unveiled September 18, 2026, AX is an Apache-2.0-licensed orchestrator/runtime for composing, coordinating, and supervising autonomous AI agents across cloud, edge, and on-device environments through a single declarative API. It exposes four Kubernetes-style primitives (Task, Workspace, Gateway, Model) under the ax.io/v1alpha1 API group, ships a Go-based CLI deployed via ko and Redis, and supports resource-efficient task suspension and resumption. It drew over 150,000 GitHub registrations within 48 hours and 600+ points on Hacker News.
Key features
Kubernetes-style declarative primitives (Task, Workspace, Gateway, Model) under ax.io/v1alpha1
Meta's non-invasive brain-to-text research model that decodes typed sentences from MEG brain scans, aimed at communication for people with speech or motor disorders.
This entry is marked for review.
Details and sources
Meta, with the Basque Center on Cognition, Brain and Language (BCBL), published Brain2Qwerty v2 on June 29-30, 2026 via the AI at Meta blog post "From Brain Waves to Words," alongside the original Brain2Qwerty v1 paper in Nature. The system uses end-to-end deep learning, fine-tuned on neural data to leverage semantic context, to decode typed sentences directly from raw magnetoencephalography (MEG) signals recorded non-invasively while volunteers type, without requiring brain surgery. Trained on roughly 22,000 sentences from nine participants (10 hours of MEG recording each), v2 reached 61% average word accuracy versus 8% for prior non-invasive approaches, with the top participant reaching 78% word accuracy and over half of decoded sentences containing one or fewer word errors; Meta released the v1/v2 training code and BCBL released the v1 dataset. This is a research system, not a shipped consumer or clinical product, aimed at eventually helping people with brain lesions or disorders that impair communication.
Key features
Non-invasive decoding via magnetoencephalography (MEG), no brain surgery required
End-to-end deep learning pipeline fine-tuned on neural data for semantic context
61% average word accuracy in v2 (vs. 8% for prior non-invasive methods); up to 78% for the best participant
Trained on ~22,000 sentences across nine participants (10 hours of MEG recording each)
Training code for v1 and v2 released by Meta; BCBL released the v1 dataset
Use cases
Assistive communication research for people with speech- or motor-impairing conditions
Academic and industry research into non-invasive brain-computer interfaces
Pricing: Not applicable — research project; training code released publicly, no commercial product or pricing.
A ChatGPT Work variant for investment banking and equity research, combining GPT-6 Astra with built-in premium financial data providers.
Details and sources
ChatGPT for Financial Services, launched September 10, 2026, is a tailored ChatGPT Work experience combining GPT-6 Astra's reasoning with built-in premium financial data from providers including Daloopa, PitchBook, LSEG News, and Crunchbase, with granular citations so bankers can trace figures back to sources. Shaped through a design partnership with Morgan Stanley and Evercore, it targets investment banking and equity research: researching companies, analyzing earnings, comparing peers, testing assumptions, and producing editable financial models, research notes, and pitchbooks. Enterprise controls include no default training on firm data, encryption at rest and in transit, and admin-configurable workspace retention.
Key features
Built-in premium data from Daloopa, PitchBook, LSEG News, and Crunchbase with granular citations
Built on GPT-6 Astra reasoning within the ChatGPT Work agent framework
Produces editable financial models, research notes, and pitchbooks
Enterprise data controls: no default training on firm data, encryption, configurable retention
Design partnership with Morgan Stanley and Evercore
Use cases
Investment banking pitchbook and research-note creation
OpenAI's agent mode in ChatGPT, powered by GPT-5.6, that gathers context across a team's tools and delivers finished sheets, slides, docs, or web apps.
Details and sources
ChatGPT Work is an agent mode in ChatGPT, powered by GPT-5.6, designed to take on more ambitious, multi-step projects: it gathers information across a team's connected apps and workflows, works independently for minutes or hours by breaking a project into smaller steps, and hands back finished materials such as spreadsheets, slide decks, reports, or working web apps. It pulls context from more than 1,400 supported plugins. OpenAI says teams at Zapier, Virgin Atlantic, and NVIDIA used ChatGPT Work to compress tasks that took weeks into hours. It is available today to all plans on desktop, and is rolling out to Plus, Pro, Business, Enterprise, and Edu on web and mobile.
Key features
Powered by GPT-5.6, works independently for minutes to hours on a single brief
Pulls context from 1,400+ connected plugins/tools
Produces finished deliverables: spreadsheets, slide decks, reports, working web apps
Breaks multi-step projects into smaller steps it completes on its own
Use cases
Turning scattered notes and drafts into finished team deliverables
Multi-hour agentic projects spanning several connected tools
Compressing multi-week analyst/ops tasks into hours
Pricing: Available today on desktop across all ChatGPT plans; rolling out to Plus, Pro, Business, Enterprise, and Edu on web and mobile at no listed separate charge beyond the underlying plan.
Anthropic's agentic coding CLI/desktop app shipped weekly updates through August-September 2026, including /design, claude plugin eval, and default auto mode.
Details and sources
From July 25 to September 23, 2026, Claude Code added: Claude Opus 5 as the default model with a 1M-token context; a Claude Security plugin running multi-agent vulnerability scans; auto mode as the default permission mode on Pro/Max/Team from August 14; a /design research preview bringing an artboard workflow into the CLI; Claude Fable 5.1 with a 1M-token context; and claude plugin eval, which scores a plugin against a no-plugin baseline.
Key features
/design artboard workflow research preview
auto mode as the default permission mode on Pro/Max/Team from August 14
claude plugin eval scores a plugin against a no-plugin baseline
Claude Opus 5 and Claude Fable 5.1 added, both with 1M-token context
Claude Security plugin runs multi-agent vulnerability scans
Use cases
Agentic coding in CLI, desktop, and IDE
Automated plugin quality testing before distribution
Vulnerability scanning integrated into the coding workflow
Pricing: Bundled into existing Pro/Max/Team/Enterprise subscription and API pricing; no separate fee.
Anthropic's latest GA frontier model for coding and knowledge work, with 75% cheaper cache reads than Fable 5.
Details and sources
Claude Fable 5.1 is Anthropic's newest generally-available frontier model, announced September 1, 2026, alongside a restricted-access twin, Claude Mythos 5.1 (the same underlying model with fewer safeguards, available only to trusted-access customers). Both run on a 1,000,000-token context window with 128,000 max output tokens. The headline change is cache-read pricing, cut 75% versus Fable 5, which Anthropic says lowers typical workload cost roughly 25% and highly agentic workload cost up to roughly 45%. Mythos 5.1 is positioned as Anthropic's most capable model for cybersecurity defense and life-sciences research (threat intelligence, vulnerability discovery, red teaming, drug discovery, and biodefense screening), scoring 52.6% on Terminal-Bench-Science and meeting or exceeding the cyber capabilities of the prior Claude Mythos 5. Mythos 5.1 shares Fable 5.1's $10/$50-per-million-token API pricing, and is reached only through Anthropic's trusted-access pathways: the Cyber Verification Program for defensive security work, and the Life Sciences Verification Program (opened September 17, 2026) for vetted biology research.
Key features
1,000,000-token context window, 128,000 max output tokens
Cache reads at $0.25/million tokens, down 75% from Fable 5
Batch processing at half price for asynchronous workloads
US-only inference option at 1.1x standard pricing
Mythos 5.1: same model, restricted trusted-access twin tuned for cybersecurity and biology, 52.6% on Terminal-Bench-Science
Mythos 5.1 access gated via the Cyber Verification Program and the Life Sciences Verification Program
Use cases
Coding and knowledge work
Long-running agents and multistep research
Document-heavy professional work
High-volume agentic workloads sensitive to cache-read cost
Pricing: $10/million input tokens, $50/million output; cache read $0.25/million (75% cut from Fable 5); batch processing $5/$25 per million; US-only inference at 1.1x standard rate.
Anthropic's near-frontier model delivering close to top-tier performance at roughly half the price.
This entry is marked for review.
Details and sources
Claude Opus 5 launched July 24, 2026, designed to match much of Claude Fable's performance on many tasks at roughly half the price. It was Anthropic's fourth Claude 5-family release within about two months of launching.
Key features
Near-Fable-tier performance at roughly half the price
Part of a rapid Claude 5-family release cadence (fourth release in under two months)
Use cases
Cost-sensitive complex workloads
General-purpose agent and coding work where Fable-tier cost isn't justified
Cloudflare's open-source, browser-based AI agent workspace giving every employee secure access to internal systems and AI tools.
Details and sources
Cloudflare announced Cloudflare OS on August 5, 2026, an open-source AI agentic workspace that runs on Cloudflare's global network. It gives employees a browser-accessible workspace equipped with AI tools and access to internal company systems without requiring new infrastructure, letting non-developers research, produce documents tied to live data, run automated workflows, and build tools without waiting on IT. It is available now as open source at github.com/cloudflare/cloudflare-os, with a managed deployment option through the Cloudflare dashboard coming later, plus implementation partners including Presidio and Happy Cog.
Key features
Browser-based, no local installation or developer expertise required
Secure per-employee workspace with governed access to internal systems
Runs on Cloudflare's global network with no new infrastructure to stand up
Fully open source, with a managed deployment option planned
Use cases
Non-technical employees running AI-assisted research and document workflows tied to live company data
Building lightweight internal tools and automations without IT involvement
Enterprise-wide AI agent rollout without per-team infrastructure work
Pricing: Free and open source (self-hosted on Cloudflare's network); a managed deployment option through the Cloudflare dashboard is planned but not yet priced.
Gives coding agents (Codex, Cursor, OpenCode, Pi) a dedicated persistent VM per task with a live app preview and PR workflow.
This entry is marked for review.
Details and sources
Cobalt, built by Amaretto Software Labs and shipped at cobaltcode.ai, is a CLI/platform that gives each coding-agent task its own dedicated virtual machine with a full repository checkout, dependencies, and running services. Users describe a task in plain language, choose an agent (Codex, Cursor, OpenCode, or Pi) to run it, watch the result run live in the browser, and open a pull request from there. The VM persists after the session ends, keeping dependencies and services running so work can resume from any device in the same conversation and app state. It integrates with repositories hosted on GitHub and Azure DevOps.
Key features
Dedicated per-task VM with its own repo checkout, dependencies, and running services
Supports multiple underlying coding agents: Codex, Cursor, OpenCode, Pi
Live, clickable app preview instead of just a code diff
Persistent state across sessions and devices
GitHub and Azure DevOps repository integration
Use cases
Running coding-agent tasks without local environment setup
Reviewing an agent's change as a running app before merging
Parallel execution of multiple coding-agent tasks across isolated VMs
Resuming an agent session and its running app from a different device
Pricing: unknown (not published in the sourced material).
Cerebras' CS-4 wafer-scale AI system, the first built on its reusable Nexus rack-scale platform, unveiled at Supernova 2026.
Details and sources
Cerebras unveiled CS-4 on August 19, 2026 at its Supernova launch event in San Francisco, the first system built on Nexus, a reusable rack-scale platform housing three Wafer-Scale Engines per rack in modular 'compute backpacks.' On August 25, 2026 at Hot Chips, Cerebras detailed the Nexus rack architecture and previewed a roadmap: CS-5 (targeting 2027, higher tokens/sec per user) and CS-6 (3D-stacked wafer-scale DRAM).
Key features
Nexus rack-scale platform hosting 3 Wafer-Scale Engines per rack
Modular 'compute backpack' design
Roadmap preview: CS-5 (2027) and CS-6 with 3D-stacked wafer-scale DRAM
Claimed throughput uplift vs. prior CS-3 generation
Cursor's coordinator-and-subagent system that runs long, multi-week engineering work like migrations to completion via a single /goal.
Details and sources
On September 10, 2026, Cursor launched Projects (beta): a non-coding coordinator agent plans and delegates work to implementer subagents running in parallel on isolated cloud VMs, sharing context across the effort. The /goal command hands the agent a long-lived objective, such as fixing all flaky tests and making CI green, pursued until complete; cloud agents keep running with the laptop closed.
Key features
Coordinator agent plans and delegates, doesn't write code itself
/goal sets a long-lived objective pursued until done
Cloud agents on isolated cloud VMs continue with the laptop closed
Parallel subagents on isolated VMs avoid conflicts
Cloud agents already generate over 60% of Cursor's own merged internal PRs
Use cases
Multi-week feature builds or migrations
Unsupervised CI/flakiness fixing
Enterprise teams delegating large backlogs to agent fleets
Pricing: unknown (beta; rolled into existing Cursor/cloud-agent usage pricing)
DeepSeek's flagship reasoning model reaches general availability with a 1M-token context and expanded agentic tool use.
Details and sources
DeepSeek-V4-Pro exited preview and became generally available on August 13, 2026 across DeepSeek's app, web interface, and API, designated DeepSeek-V4-Pro-0813. It activates 49B parameters per token, supports a 1-million-token context window with up to 384,000-token output, and expanded thinking-effort settings to low/high/max, focused on agentic tasks such as tool use, code execution, and multi-step workflows.
Key features
49B active parameters (Mixture-of-Experts)
1M-token context window, 384K max output
Three thinking-effort levels: low, high, max
MIT license
Use cases
Large coding projects
Mathematical reasoning
Long-document analysis
Autonomous multi-step agent workflows
Pricing: MIT license, weights free; API pricing increased 2026-08-16 16:00 UTC (exact new rate unknown).
DeepSeek's 552B-parameter MoE model with native image+text input, a 1M-token context window, and MIT license.
Details and sources
DeepSeek-V4.1-Flash is a 552-billion-parameter Mixture-of-Experts model (8B active on prefill, 16B on decode) released September 10, 2026, reading images and text natively with context up to 1 million tokens and up to 384,000 output tokens. It uses a Compressed Expert Dispatch (CED) architecture and is published on Hugging Face under the MIT license with an accompanying technical report.
Key features
552B total parameters, 8B/16B active (prefill/decode)
1M-token context, 384K max output
Native text + image input
Compressed Expert Dispatch (CED) architecture
MIT license
Use cases
Cost-sensitive multimodal agentic tasks
High-volume API workloads (peak/off-peak pricing)
Pricing: MIT license, weights free; API peak $0.30/M input (cache miss), $0.006/M (cache hit), $1.20/M output; off-peak half those rates.
Infrastructure and inference/servinginfrastructure
DigitalOcean
DigitalOcean Managed Agents
DigitalOcean's managed service combining isolated agent execution, governed tool access, and serverless inference in one cloud offering.
Details and sources
DigitalOcean launched the public preview of Managed Agents on September 22, 2026, unifying agent execution, tool access, and inference on one cloud stack. Each agent session runs in its own hardware-isolated harness runtime (compatible with harnesses teams already use), with governed access to more than 16,000 tools and Serverless Inference across 75-plus open and proprietary models, metered only while the agent is active. Launch partners building on the platform include OpenHands, Qencode, and Amplitude.
Key features
Hardware-isolated harness runtime per agent session with separate credentials management
Governed access to 16,000+ tools
Serverless Inference across 75+ open and proprietary models
Usage-based metering that only runs while an agent is active
Use cases
Running production agentic workloads without building custom sandboxing/isolation infrastructure
Giving agents governed, auditable access to a large tool catalog
Combining inference and agent execution billing on a single cloud
Pricing: Public preview; usage-based metering that runs only while an agent is active (exact rates not yet published at preview launch).
Infrastructure and inference/servinginfrastructure
Equinix
Equinix Inference Exchange
Equinix announced a distributed AI inference program with NVIDIA and Together AI supporting 200+ open models, availability Q1 2027.
This entry is marked for review.
Details and sources
On September 2, 2026, Equinix announced Equinix Inference Exchange, combining NVIDIA's validated Enterprise Reference Architectures with Together AI's inference platform (200+ open-source models) for metro-edge inference, open-model migration, and sovereign AI use cases. The program itself becomes available starting Q1 2027 -- this is a partnership announcement, not yet a shipped product.
CrowdStrike's AI Detection and Response product that discovers, monitors, and enforces policy on agentic AI running inside an organization.
Details and sources
Unveiled at Fal.Con 2026 (Las Vegas, Aug 31-Sept 3) on September 1, 2026, Falcon Guardian gives security teams discovery of an organization's full AI agent fleet, including unapproved 'shadow AI', runtime visibility into what agents access and do, and policy enforcement to stop threats before they act. It is the expanded successor to Falcon AI Detection and Response, which reached general availability in December 2025.
Key features
Fleet-wide discovery of sanctioned and shadow AI agents
Runtime monitoring of agent actions and resource access
Policy-to-enforcement pipeline to block agent-driven threats
Successor to Falcon AI Detection and Response
Use cases
Finding unapproved AI agents deployed by employees/teams
Runtime guardrails for enterprise agentic AI deployments
SOC visibility into AI-agent attack surface
Pricing: Not disclosed; part of the Falcon platform.
Multimodal flow model unifying video, audio and (soon) image generation, priced per second, with a robotics offshoot running on Audi production lines.
Details and sources
Black Forest Labs unveiled FLUX 3 on July 23, 2026, a unified Self-Flow architecture model jointly trained on image, video, and audio, extensible to action prediction for robotics. At launch, video generation (up to 20 seconds, 4K, native audio, keyframes/continuation) was live; image modality was announced as coming soon, and an open-weight FLUX 3 Dev edition is planned later in 2026. A derivative, FLUX-mimic, is already running on Audi production lines.
Key features
Unified Self-Flow architecture spanning image, video, and audio
Up to 20-second video clips at up to 4K, with native audio
Per-second pricing: $0.17/s (HD) to $0.80/s (4K)
Robotics action-prediction head (FLUX-mimic, deployed at Audi)
Open-weight FLUX 3 Dev planned later in 2026
Use cases
Short-form AI video with native audio
Robotics action prediction and manufacturing QA
Future open-weight local image/video generation
Pricing: Per-second video pricing: $0.17/s at HD up to $0.80/s at 4K; image pricing not yet published.
Google's specialized Gemini 3.5 Flash variant fine-tuned to find and patch software vulnerabilities via CodeMender.
Details and sources
Gemini 3.5 Flash Cyber launched July 21-22, 2026 as a cybersecurity-specialized fine-tune of Gemini 3.5 Flash, built to detect, validate, and patch vulnerabilities across large codebases and operate inside Google's CodeMender agent infrastructure. In testing against the V8 JavaScript engine, it found 55 unique confirmed vulnerabilities versus 47 for base Gemini 3.5 Flash and 36 for Claude Opus 4.6, including 10 issues both competitors missed. It is currently a limited-access pilot for governments and trusted partners rather than a public API model.
Key features
Fine-tuned specifically for vulnerability discovery and patching
Integrated with Google's CodeMender security agent
Found 55 unique confirmed vulnerabilities in V8 testing vs. 47 (Gemini 3.5 Flash) and 36 (Claude Opus 4.6)
Competitive performance at the frontier on the CyberGym benchmark
Lower cost per token than larger frontier models
Use cases
Large-scale automated vulnerability hunting across codebases
Continuous repository security scanning
Automated security patch generation
Pricing: Not publicly priced; limited-access pilot program for governments and trusted partners via CodeMender, not generally available.
Google's fastest, cheapest Gemini 3 model for high-throughput agentic search and document processing.
Details and sources
Gemini 3.5 Flash-Lite launched July 21, 2026 alongside Gemini 3.6 Flash and Gemini 3.5 Flash Cyber, positioned as the high-throughput, low-latency tier of the Gemini 3 family for agentic search and data-intensive tasks. It generates up to 350 output tokens per second, the fastest in the 3.5 series, and outperforms Gemini 3 Flash on some benchmarks (SWE-Bench Pro: 54.2% vs. 49.6%). It remains Google's cheapest current-generation model, priced at $0.30/M input and $2.50/M output tokens, and as of September 2026 has not been superseded by a later Flash-Lite release.
Key features
350 output tokens/second, the fastest model in the Gemini 3.5 series
Built-in computer use tool for agentic workflows
SWE-Bench Pro: 54.2% (vs. 49.6% for Gemini 3 Flash)
Significant quality gains over predecessor Gemini 3.1 Flash-Lite
Google's 1M-context workhorse Gemini model, faster and cheaper than 3.5 Flash, launched alongside 3.5 Flash-Lite and 3.5 Flash Cyber.
Details and sources
Gemini 3.6 Flash launched July 21, 2026 alongside Gemini 3.5 Flash-Lite and Gemini 3.5 Flash Cyber, priced at $1.50/million input and $7.50/million output tokens. It keeps a 1,048,576-token context window with a 64,000-token output cap, moves the knowledge cutoff to March 2026, accepts text, image, video, audio, and PDF input, and uses about 17% fewer output tokens than 3.5 Flash while scoring higher on coding, long-context, and computer-use benchmarks at roughly 280 tokens/second. It was superseded three weeks later by Gemini 3.7 Flash.
Key features
1,048,576-token context window, 64,000-token max output
Knowledge cutoff moved forward to March 2026
Multimodal input: text, image, video, audio, PDF
~17% fewer output tokens than 3.5 Flash at higher coding/long-context/computer-use scores
~280 tokens/second generation speed
Use cases
General-purpose coding and agentic workloads needing large context
Computer-use tasks
Cost-sensitive production traffic prior to the 3.7 Flash upgrade
Google's most intelligent workhorse Gemini model yet for coding and agents, with introductory $0.75/M input pricing.
Details and sources
Gemini 3.7 Flash launched August 13, 2026, three weeks after Gemini 3.6 Flash, as Google's most capable workhorse model for coding and agentic workflows to date. It has a 1,048,576-token context window and carries introductory pricing of $0.75 per million input tokens through December 31, 2026.
Key features
1,048,576-token context window
Introductory pricing of $0.75/million input tokens through 2026-12-31
Marketed improvements across software engineering, web development, and agentic workflows
Use cases
Coding and agentic workflows
Software engineering and web development tasks
Pricing: $0.75/million input tokens (introductory, through 2026-12-31); output pricing unknown.
Google's cost-efficient real-time voice model for natural conversation with live visual context and background tasks.
Details and sources
Gemini 3.8 Live launched September 15, 2026 as a cost-efficient model built for real-time voice conversations with live visual input and the ability to run tools or API calls in the background without interrupting dialogue. It automatically detects and switches between 97 languages mid-conversation and is aimed at production voice assistant use cases across consumer and enterprise surfaces.
Key features
Real-time, near-instant voice conversation with live visual context
Automatic detection and mid-conversation switching across 97 languages
Background tool and API execution without interrupting the conversation
Positioned as the cost-efficient tier versus Gemini 3.8 Live Extended Thinking
Use cases
Voice assistants and conversational agents
Real-time customer support with visual context
Multilingual voice interfaces
Pricing: Not disclosed in the launch announcement.
Google's real-time voice model that reasons and speaks simultaneously for complex multi-step conversational tasks.
Details and sources
Gemini 3.8 Live Extended Thinking launched September 15, 2026 alongside Gemini 3.8 Live, built for complex reasoning tasks that require sustained thinking while keeping the conversation flowing in real time. Google states it reasons and speaks simultaneously while handling multi-step tasks, and it topped Artificial Analysis' Speech to Speech Quality Index with a score of 82.6 alongside high agentic task-completion scores.
Key features
Simultaneous reasoning and speech during live conversation
Top score (82.6) on Artificial Analysis' Speech to Speech Quality Index at launch
High agentic task-completion performance in voice-driven multi-step tasks
Automatic detection and mid-conversation switching across 97 languages, inherited from the Live line
A NotebookLM-style notebook feature built directly into the Gemini app, rolling out to Workspace and personal accounts.
Details and sources
Google began gradually rolling out a new Notebooks feature inside the Gemini app on September 14, 2026, available to Workspace customers, Workspace Individual subscribers outside the EEA, and personal Google account users globally. Users open a side panel, start a New notebook, and add up to 10 sources with a defined focus area.
Key features
In-app notebook creation from the Gemini side panel
Up to 10 sources per notebook
Rolling out to personal and Workspace accounts (EEA excluded initially)
Use cases
Source-grounded research and note-taking inside Gemini
Study/document synthesis without leaving the chat app
Pricing: unknown -- feature appears bundled into existing Gemini app access tiers, no separate charge disclosed.
Google DeepMind's embodied-reasoning model that acts as a high-level planning brain for robots, with multi-robot collaboration.
Details and sources
Gemini Robotics ER 2 launched July 30, 2026 alongside Gemini Robotics 2 and Gemini Robotics On-Device 2, serving as a high-level embodied-reasoning brain that orchestrates lower-level vision-language-action models. It adds whole-body humanoid control, continuous video-based progress tracking, and multi-robot collaboration over its predecessor, Gemini Robotics-ER 1.5. Google reports 91.3% moment-finding accuracy (0.96s mean absolute distance) at roughly four times the execution speed of the prior generation.
Key features
Multi-robot collaboration for shared-space, multi-step workflows
Continuous video feed analysis for real-time progress tracking and self-correction
91.3% moment-finding accuracy, 0.96s mean absolute distance, ~4x execution speed vs. prior generation
57.4% progress-classification accuracy
Low-latency execution via the Gemini Live API; improved human-proximity safety detection
Use cases
Humanoid and multi-robot task planning and coordination
Long-horizon, multi-step physical task execution
Robotics developers building on vision-language-action stacks
Pricing: Not disclosed; available via Gemini API and Google AI Studio, with private preview on Gemini Enterprise Agent Platform.
Zhipu's flagship coding/reasoning model with a 1M-token context, open-weighted about two weeks after launch.
This entry is marked for review.
Details and sources
Zhipu AI (operating internationally as Z.ai) released GLM-5.3 on August 14, 2026, built via scaled post-training on the GLM-5 architecture for coding, reasoning, and agentic tool-use workloads, under a custom license requiring hosts with over $10B in trailing-12-month revenue to pass a Z.ai security review. Zhipu open-weighted a GLM-5.3-Flash variant on Hugging Face (zai-org) around August 26, 2026, under the MIT license.
Key features
1M-token context window
MIT-licensed GLM-5.3-Flash weights (released ~2 weeks after launch)
Tied with Kimi K3 on the Artificial Analysis Intelligence Index at a score of 60
Flash variant for lower-cost deployment
Use cases
Agentic coding workloads
Tool-driven reasoning tasks
Pricing: MIT license for GLM-5.3-Flash weights; hosted pricing via Z.ai's GLM Coding Plan, exact rate unknown.
Google Cloud Developer Plugin for AI Coding Agents
An installable plugin that equips AI coding agents (Claude Code, Codex CLI, Antigravity) with Google Cloud skills, guardrails, and docs access.
Details and sources
Announced September 10, 2026, the Google Cloud Developer Plugin is an installable bundle that gives AI coding agents environment-aware skills for working with Google Cloud, built on the open, vendor-neutral Agent Plugins specification for packaging Agent Skills and MCP servers. It runs background checks for CLI prerequisites and existing projects, reviews workflows for IAM/security risks such as accidental key leaks, and bundles the Developer Knowledge MCP server for up-to-date access to official Google Cloud documentation. It installs into Google's Antigravity CLI, Claude Code (via the google/skills marketplace), and Codex CLI.
Key features
Environment-awareness checks for gcloud CLI availability and existing project configuration
IAM best-practice review to catch risks like accidental key leaks or unsafe git commits
Bundled Developer Knowledge MCP server for live access to official Google Cloud docs
Installable in Antigravity CLI, Claude Code, and Codex CLI via the open Agent Plugins spec
gcloud CLI guardrails for safer resource interactions
Use cases
Giving AI coding agents safe, guided access to Google Cloud projects
Preventing credential/IAM mistakes made by autonomous coding agents
Keeping coding-agent responses grounded in current Google Cloud documentation
Pricing: Free; distributed via the open Google Agent Skills repository (g.dev/cloud/agent-plugins). Requires a DEVELOPERKNOWLEDGE_API_KEY for the bundled MCP server.
OpenAI's mid-tier GPT-5.6 model, priced about half of GPT-5.5 with comparable everyday capability.
Details and sources
GPT-5.6 Terra is the middle tier of OpenAI's three-model GPT-5.6 family (Sol, Terra, Luna), announced June 26, 2026 in a limited preview to a small group of trusted partners and reaching general availability across ChatGPT, Codex, and the API on July 9, 2026. Terra is positioned as competitive with GPT-5.5 while costing roughly half as much, at $2.50/million input tokens and $15/million output tokens. Under OpenAI's Preparedness Framework, Terra is classified High capability in both cybersecurity and biological/chemical risk, alongside Sol and Luna. Unlike Sol and Luna, Terra was not carried forward into OpenAI's September 2026 GPT-6 refresh, which shipped GPT-6 Sol and GPT-6 Luna but no GPT-6-generation Terra tier.
Key features
Mid-tier pricing and capability between flagship Sol and budget Luna
Roughly 2x cheaper than GPT-5.5 at comparable everyday capability
Classified High capability under OpenAI's Preparedness Framework for cyber and bio/chem risk
Available via ChatGPT, Codex, and the API since general availability
Use cases
Everyday reasoning and writing tasks at lower cost than a flagship-tier model
Cost-sensitive production workloads that don't need Sol-level reasoning depth
OpenAI's flagship model for computer/browser use, coding, and cybersecurity, rolling out via ChatGPT and API.
Details and sources
GPT-6 Astra is OpenAI's flagship model, launched September 3, 2026, positioned above GPT-6 Sol and GPT-6 Luna. It has a 1,050,000-token context window with 128,000 max output tokens, and OpenAI markets it as state-of-the-art for computer/browser use, software engineering, deep research, and cybersecurity. It rolled out first to companies in OpenAI's Daybreak cybersecurity program, then to ChatGPT Plus/Pro/Business/Enterprise and the API and AWS.
Key features
1,050,000-token context window, 128,000 max output tokens
Positioned as state-of-the-art for computer/browser-use agentic tasks
Phased rollout starting with Daybreak cybersecurity program participants
Available via ChatGPT plans, the API, and AWS
Use cases
Advanced software engineering and deep research
Computer/browser-use agentic tasks
Cybersecurity work under the Daybreak program
Scientific and document-heavy professional work
Pricing: $10/million input tokens, $50/million output, $1/million cache read, $12.50/million cache write, $10 per 1,000 web-search calls below 272K input tokens; rates double for input/cache and rise 1.5x for output above that threshold.
OpenAI's most efficient GPT-6 model for high-volume, low-cost tasks like summarizing, extracting, and quick Q&A.
Details and sources
GPT-6 Luna is OpenAI's cheapest GPT-6 model, positioned for high-volume tasks with a clear goal such as summarizing documents, extracting information, or answering quick questions. It shares GPT-6 Sol's 1,050,000-token context window and 128,000-token max output, but with a May 18, 2026 knowledge cutoff and pricing an order of magnitude lower. It launched alongside GPT-6 Sol on September 22, 2026, available in ChatGPT Work, Codex, and the API for most paid accounts, plus the desktop app and Free/Go tiers.
Key features
1,050,000-token context window, 128,000 max output tokens
Lowest-cost model in the GPT-6 lineup at launch
Supports streaming, structured outputs, function calling, file search, image input, web search, and prompt caching
Responses API tool support: web search, file search, image generation, code interpreter, computer use, hosted shell
Available to Free and Go ChatGPT tiers, not just paid plans
Use cases
High-volume document summarization and information extraction
Quick question-answering at low per-token cost
Free-tier and desktop-app ChatGPT usage
Lightweight agentic steps where Sol/Astra-level reasoning isn't needed
OpenAI's mid-tier GPT-6 model for complex coding and agentic workflows, priced at half its GPT-5.6 predecessor.
Details and sources
GPT-6 Sol is OpenAI's balanced model for interactive and agentic coding, sitting below GPT-6 Astra in price and capability. It ships with a 1,050,000-token context window (922,000 max input, 128,000 max output), a April 20, 2026 knowledge cutoff, and configurable reasoning effort. OpenAI says Sol makes about half as many mistakes as GPT-5.6 Sol, reaching Astra-level reliability at lower cost, and prices it 50% below GPT-5.6 Sol's promotional rate. It became available via the API on September 22, 2026, with a gradual rollout to ChatGPT Work and Codex for eligible paid plans.
Key features
1,050,000-token context window, 128,000 max output tokens
Configurable reasoning effort from none through max
Supports streaming, structured outputs, function calling, file search, image input, web search, and prompt caching
Responses API tool support: web search, file search, image generation, code interpreter, computer use, hosted shell, patch application, MCP
Requests over 272K input tokens billed at 2x input/cache and 1.5x output rates
Use cases
Interactive and agentic coding tasks
Multistep validation workflows that need careful, sustained reasoning
Codex and GitHub Copilot coding sessions
Cost-sensitive agentic tasks that previously required Astra-tier pricing
OpenAI's full-duplex voice model that listens and speaks simultaneously, now available in the API at $0.05/minute.
Details and sources
GPT-Live-1 brings ChatGPT's natural, full-duplex voice conversation model to the API, launched September 10, 2026 at $0.05 per minute for the voice front-end layer. It listens and speaks at the same time, handles interruptions, pauses, and backchannels, and delegates deeper reasoning and actions to whichever model or tools it is paired with, including telephony support for full-duplex phone agents. OpenAI began rolling GPT-Live-1 and the smaller GPT-Live-1 Mini out to ChatGPT globally on July 8, 2026, with GPT-Live-1 becoming the default ChatGPT Voice model for Go, Plus, and Pro users and GPT-Live-1 Mini serving Free users.
Key features
Full-duplex: listens and speaks simultaneously, handling interruptions and backchannels
Silent context management without narrating every step
Context retention across extended sessions and telephony support for phone agents
Delegates reasoning/actions to paired models and tools
GPT-Live-1 Mini variant serves ChatGPT Free users
Use cases
Natural voice assistants and customer-facing phone agents
Full-duplex conversational interfaces in third-party apps via the API
Default ChatGPT Voice experience across paid and free tiers
Pricing: $0.05/minute for the GPT-Live-1 voice layer via the API; included as the default voice model in ChatGPT Go/Plus/Pro (GPT-Live-1 Mini for Free tier).
OpenAI's frontier reasoning model built for biology, drug discovery, and translational medicine research, named after Rosalind Franklin.
Details and sources
GPT-Rosalind is OpenAI's frontier reasoning model purpose-built to accelerate drug discovery, genomics analysis, protein reasoning, and translational medicine research, with improved tool use across chemistry, protein engineering, and genomics. It is available in ChatGPT, Codex, and the API for qualified customers through OpenAI's trusted access program, and OpenAI is working with life-sciences partners including Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific to apply it across research workflows. The model is named after Rosalind Franklin, whose research contributed to revealing the structure of DNA. OpenAI later updated the series to GPT-Rosalind-5.5, incrementally trained from GPT-5.5 with stronger medicinal chemistry and genomics performance, still gated to trusted research organizations.
Key features
Frontier reasoning tuned for biology, drug discovery, and translational medicine
Improved tool use and understanding across chemistry, protein engineering, and genomics
Available through OpenAI's trusted access program, not general self-serve access
Launch partners include Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific
Use cases
Target discovery and hypothesis generation in drug development
Genomics and protein-engineering analysis workflows
Translational medicine research at pharma and biotech organizations
Pricing: Gated trusted-access program for qualified life-sciences customers; pricing not publicly listed.
IBM's open-weight enterprise reasoning model family in 3B/8B/30B sizes with native chain-of-thought and tool calling.
Details and sources
IBM released Granite 4.2 on August 25, 2026, a family of open-weight reasoning language models in 3B, 8B, and 30B parameter sizes under Apache 2.0 with 128K context windows. It is IBM's first Granite generation to natively combine chain-of-thought reasoning with tool calling, exposing a thinking/non-thinking switch plus a low-effort mode for simple queries.
Key features
3B, 8B, and 30B parameter sizes
128K context window
Apache 2.0 license
Native chain-of-thought plus tool calling
Thinking/non-thinking toggle with a low-effort mode
Use cases
Enterprise agentic workflows
Cost-tiered reasoning (toggle effort by task difficulty)
xAI's flagship reasoning model for coding, agentic, and engineering tasks, now trained partly on SpaceX/Starlink engineering data.
Details and sources
Grok 4.7 is xAI's flagship model, released September 21, 2026, at 2.1 trillion parameters (up 40% from Grok 4.6's 1.5 trillion) at the same price and speed as its predecessor. xAI trained it with a longer reinforcement-learning run on harder, longer-horizon tasks and folded in supplemental data from SpaceX -- Starlink telemetry, manufacturing records, and engineering failure logs -- aimed at improving reasoning about hardware and physical systems. It has a 500,000-token context window with text and image input, text-only output, and no output length limit, and is available now in Cursor, Grok Build, the Grok API, GitHub Copilot (Pro, Pro+, Max, Business, Enterprise), and third-party coding harnesses.
Key features
2.1 trillion parameters, up from 1.5 trillion in Grok 4.6
500,000-token context window, text and image input, text-only output with no output length limit
Configurable reasoning effort: low, medium, high (default), xhigh
Trained in part on SpaceX/Starlink telemetry, manufacturing, and engineering failure-log data
Grok 4.7 Fast variant available at 2x token rates in Cursor and Grok Build only
Use cases
Agentic coding in Cursor, Grok Build, and GitHub Copilot
Long-horizon engineering and hardware-reasoning tasks
General agentic and office-work automation
Third-party coding harnesses and model routers via the Grok API
Pricing: $2/million input tokens, $0.50/million cached input, $6/million output below 200K prompt tokens; $4/$1/$12 per million above 200K. Grok 4.7 Fast billed at 2x standard rates.
xAI's autonomous 'AI teammate' agents that run persistently in the cloud and now ship with more SuperGrok and Cursor plans.
Details and sources
Grok Bot launched in beta on August 11, 2026 as an autonomous agent designed to be handed an entire project or function rather than a single prompt, running in a persistent cloud computer so work continues while the user is away. As of September 2026, xAI expanded access so Grok Bot is included with SuperGrok, SuperGrok Plus, SuperGrok Heavy, Cursor Pro+, Cursor Ultra, and both Cursor Teams tiers, with its own usage allowance separate from existing Grok and Cursor quotas. It runs on macOS, Windows, Linux, iOS, and Android.
Key features
Persistent cloud-hosted 'teammate' agents that keep working when the user is offline
Separate usage allowance from the underlying Grok/Cursor subscription
Bundled into SuperGrok Plus/Heavy and Cursor Pro+/Ultra/Teams plans at no extra cost
Use cases
Office-manager style agents that intake work orders and plan capacity across tools
Inbox-management agents that triage and clear email
Handing off multi-step projects or an entire function to an agent instead of prompting step by step
Pricing: Included at no added cost with SuperGrok Plus, SuperGrok Heavy, Cursor Pro+, Cursor Ultra, Cursor Teams Standard, and Cursor Teams Premium; usage is metered separately from the base plan.
xAI's open-source terminal coding agent that reads and edits codebases, runs commands, and spawns up to eight subagents.
Details and sources
Grok Build is xAI's terminal-based coding agent harness and TUI, introduced in July 2026 as an early beta limited to SuperGrok Heavy. It left beta at v1.0 on August 7, 2026 and is now open source under Apache 2.0 (github.com/xai-org/grok-build), reading a codebase, editing files, running commands, and spawning up to eight subagents in isolated git worktrees. As of September 2026 it is available on every plan on web and mobile, and a fast standalone model, Grok Build 0.1, entered public beta on the xAI API.
Key features
Terminal CLI plus fullscreen, mouse-interactive TUI
Spawns up to eight subagents in isolated git worktrees
Open source under Apache 2.0
Publishing/sharing built apps, X integration, and access to xAI's own models from built apps
Grok Build 0.1 standalone fast coding model available via the xAI API in public beta
Use cases
Autonomous multi-file coding and refactoring from the terminal
Parallel agentic coding tasks split across isolated git worktrees
General-purpose building beyond software engineering (e.g. games, simulations)
Pricing: Included on every xAI/Grok plan (previously limited to SuperGrok Heavy beta); Grok Build 0.1 model available via xAI API in public beta, pricing not separately published as of this writing.
Legal AI platform ships Agentic Vault Search, Horizon Scanning, and Harvey Academy; raises $550M at a $15.5B valuation.
Details and sources
In September 2026, Harvey introduced Agentic Vault Search and Organization (general availability for search, early access for organization) for large document Vaults, Horizon Scanning (early access) for tracking regulatory and legislative change with custom alerts, and Harvey Academy for on-demand training. Separately, on September 9, 2026, Harvey raised $550M at a $15.5-15.6B valuation. It also launched a pro bono program in the UK, Australia, and US for legal-aid casework.
Key features
Agentic Vault Search (GA) and Organization (early access)
Horizon Scanning: regulatory/legislative change tracking with custom alerts
Harvey Academy: on-demand training and workflows
Pro bono program for housing/veterans/social-welfare casework
Use cases
Large-scale legal document search and organization
Regulatory change monitoring for compliance teams
Legal team onboarding and training
Pro bono legal-aid agent workflows
Pricing: unknown (enterprise contracts, not published).
Hebbia launches Max, a dealmaking AI agent, and Matrix 2.0, an agent that takes actions across financial workflows.
Details and sources
Hebbia launched Max, a dealmaking AI agent, on August 14, 2026, followed by Matrix 2.0 on August 26, 2026 -- a rebuilt version of its flagship product that can take actions rather than only answer queries. The launches came alongside expanded data partnerships (Snowflake, Fitch, Intercontinental Exchange, Intralinks) feeding structured and market data directly into the platform.
Key features
Max: autonomous dealmaking agent
Matrix 2.0: action-taking agent workflows (not just Q&A)
Snowflake integration for structured data
Fitch ratings/rationale data access
Intercontinental Exchange near-real-time equity/ETF data
Thinking Machines Lab's open-weight multimodal MoE model with a 1M-token context and native text, image, and audio input.
Details and sources
Inkling is an open-weight multimodal large language model released by Thinking Machines Lab on July 15, 2026, built as a decoder-only mixture-of-experts model with 975B total and 41B active parameters (a smaller Inkling-Small variant has 276B total and 12B active parameters). It natively processes text, image, and audio inputs, was trained on roughly 45 trillion tokens spanning text, images, audio, and video, and supports a 1-million-token context window with variable reasoning-effort levels from "none" to "max."
Key features
Decoder-only MoE: 975B total / 41B active parameters (Inkling-Small: 276B total / 12B active)
Native multimodal input: text, image, and audio
1M-token context window with hybrid global/sliding-window attention
Variable reasoning-effort levels and multi-token prediction layers for speculative decoding
Available in BF16 and quantized NVFP4 variants
Use cases
Long-context multimodal reasoning over mixed text, image, and audio inputs
Self-hosted deployment on GPU clusters via Inference Endpoints
Real-time voice and multimodal application prototyping
Pricing: Open-weight model, free to download; Hugging Face Inference Endpoints deployment on 8x RTX PRO 6000 quoted at $22/hour of uptime.
Moonshot AI's preview multimodal upgrade to its Kimi K2 open-weight agentic model line, released September 2026.
This entry is marked for review.
Details and sources
Kimi K2.8 Preview was released September 11, 2026, the latest in Moonshot AI's Kimi K2 series (K2.5 in January 2026, K2.6 in April 2026, K2.7-Code in June 2026), continuing the line's rapid cadence and multimodal, agentic focus.
Key features
Multimodal preview release
Continues the Kimi K2 series' rapid release cadence
Continues Agent Swarm multi-agent coordination lineage from K2.5/K2.6
Moonshot's 2.8-trillion-parameter model, the largest open weights ever released, under a custom revenue-gated license.
Details and sources
Moonshot AI announced Kimi K3 on July 16, 2026, with weights following on July 27, 2026, at 2.8 trillion parameters -- the largest open-weights model released to date. It debuted at #3 on the Artificial Analysis leaderboard behind Claude Fable 5 and GPT-5.6 Sol, but led on Arena.ai's front-end web development benchmark. Weights are released under a custom license requiring companies with more than $20M in annual revenue to negotiate a contract with Moonshot before offering K3 to external customers as a service.
Key features
2.8 trillion parameters (largest open-weights model to date)
#3 on the Artificial Analysis leaderboard at launch
Leads Arena.ai's front-end web development benchmark
Langfuse v4 went GA for self-hosted deployments with 10x faster dashboards and new monitors/code evaluators.
Details and sources
Langfuse v4 became generally available for self-hosted deployments on August 17, 2026. It adds monitors for cost/quality/latency thresholds, code evaluators for deterministic checks, full-text search, a 'Pulse' metric visualization, a natural-language 'Langfuse Assistant', and Observations/Metrics API v2 -- cutting initial large-dataset table loads from seconds to milliseconds and making long-range dashboards load over 10x faster. Self-hosted v3 gets security patches through January 2027; Langfuse Cloud's forced v4 cutover is November 16, 2026.
Key features
Cost/quality/latency threshold monitors
Deterministic code evaluators
Full-text search across traces
Langfuse Assistant (natural-language queries)
Observations API v2 / Metrics API v2
10x+ faster dashboard loads at scale
Use cases
Self-hosted LLM observability at large trace volumes
Automated quality/cost/latency alerting
Natural-language exploration of trace data
Pricing: Open-source self-hosted (free) and Langfuse Cloud tiers; no new pricing announced with v4.
An open-weight, self-hostable typed-decision model positioned as the open-source alternative to TypeSafe AI's Jev.
Details and sources
Laya, released September 18, 2026 under Apache 2.0, is a 421-million-parameter decision model from Convai Innovations built on a ModernBERT-large backbone with a from-scratch decision head (2 transformer layers, an option-marker scorer, and an act/escalate head). It answers typed questions (choice, score, boolean) with calibrated probabilities instead of generating text, positioned as the open-weight, self-hostable counterpart to TypeSafe AI's managed Jev service, which launched three days earlier. On Convai's typed-decisions benchmark, Laya scores 0.766 accuracy and 0.081 ECE calibration error versus Jev's 0.727 accuracy and 0.246 ECE, and responds in 32.8ms on a T4 GPU versus Jev's 236-276ms, though it falls behind Jev on high-cardinality questions (50+ options): 0.425 versus Jev's 0.870 on a 72-label test.
Key features
421M parameters, ModernBERT-large backbone plus a from-scratch decision head, Apache 2.0 licensed
32.8ms response time on a T4 GPU, about 7.8x faster than Jev's 236-276ms
Higher accuracy (0.766 vs 0.727) and calibration (0.081 vs 0.246 ECE) than Jev on typed-decision benchmarks
Multilingual variant (laya-multilingual, mmBERT-base) covers 100+ languages; English checkpoint covers 23 with routing to 45+
Runs in-browser via Hugging Face Spaces demo, via `pip install laya`, or self-hosted with an HTTP server compatible with existing TypeSafe clients
Use cases
Self-hosted routing, classification, and scoring where data cannot leave the local environment
Fine-tuning or pinning a specific decision-model version for private data
Low-latency typed decisions in-browser with no server or API key
Drop-in replacement for TypeSafe Jev clients via a compatible HTTP server
Pricing: Free and open source (Apache 2.0); no API pricing, self-hosted deployment costs only your own compute.
Mistral's open-weight Lean 4 theorem-proving model that solved 587 of 672 PutnamBench problems and saturated miniF2F.
Details and sources
Leanstral 1.5, released July 2, 2026 under Apache 2.0, is a 119B-total/6B-active-parameter model specialized for writing and checking formal proofs in Lean 4. Mistral reports 100% on the miniF2F benchmark, 587 of 672 solved problems on PutnamBench, and top open-source scores of 87% on FATE-H and 34% on FATE-X, with the model already surfacing five previously unknown bugs while scanning 57 open-source repositories. It supports a 256k-token context, is available free via a Mistral API endpoint and Hugging Face weights, and is integrated into Mistral Vibe.
Key features
119B total / 6B active parameters, Apache 2.0 licensed
100% on miniF2F; 587/672 on PutnamBench; 87% FATE-H / 34% FATE-X
256k-token context with strong test-time scaling up to 4M-token budgets
Agentic code verification; found 5 unknown bugs across 57 scanned repositories
Free API endpoint (leanstral-1-5) and integration with Mistral Vibe
Use cases
Formal verification of mathematical proofs in Lean 4
Automated bug discovery via code-to-proof verification
Competition-math and research-level theorem proving
Pricing: Free: Apache 2.0 weights on Hugging Face and a free API endpoint (leanstral-1-5).
Hugging Face's open-source robotics toolkit adds world-model policies, reward models, and simulation benchmarks to close the robot learning loop.
Details and sources
Hugging Face released LeRobot v0.6.0 on July 7, 2026, extending its open-source embodied-AI toolkit from data collection into a full imagine-evaluate-improve loop. The release adds world-model policies (VLA-JEPA, FastWAM, LingBot-VA), more vision-language-action models, reward-model support, six simulation benchmarks under lerobot-eval, a rollout CLI that turns failures into training data, FSDP training, and HF Jobs cloud training, plus depth sensing and VLM-powered dataset annotation. It also ships breaking changes: `pip install lerobot` no longer bundles dataset/training dependencies by default, and some internal import paths were replaced with canonical public entry points.
Liquid AI's open-weight LFM2.5-Encoder-230M and -350M models, bidirectional encoders tuned for fast long-context CPU inference.
Details and sources
Liquid AI released LFM2.5-Encoder-230M and LFM2.5-Encoder-350M on July 28, 2026, open-weight bidirectional encoder models converted from its LFM decoder architecture via attention-mask modification and symmetric convolution padding. The models support up to 8,192-token context with gradually increasing latency, match or beat larger encoders on GLUE, SuperGLUE, and multilingual benchmarks, and run roughly 3.7x faster than ModernBERT-base on CPU at maximum context.
Key features
Two sizes: 230M and 350M parameters, open-weight
Bidirectional encoders derived from decoder architecture via attention-mask and convolution-padding changes
8,192-token context support with gradual latency scaling
~3.7x faster than ModernBERT-base on CPU at max context
Supports masked language modeling for fine-tuning
Use cases
Text classification and routing on CPU-only infrastructure
PII detection and policy linting over long documents
Prompt routing and spell-checking pipelines
Pricing: Free and open weight; hosted on Hugging Face.
Vetted-access program giving life-science researchers Claude Mythos, Opus, and Sonnet with more permissive biology safeguards.
Details and sources
Anthropic opened applications for the Life Sciences Verification Program on September 17, 2026, moving what had been a narrow early-access cohort into a broader beta for qualifying teams and institutions. The program gives verified life-science professionals access to Claude Mythos, Opus, and Sonnet with adjusted biology-related safeguards for work such as drug discovery, research biology, clinical development, and manufacturing that is otherwise blocked in generally available models; non-life-sciences safeguards stay in place. Access has two tiers: Standard Use, which covers most routine life-sciences work, extends to a team, and renews annually; and High-risk Use, an add-on for a single named project that renews every six months and, for now, applies to Opus 5 and Sonnet 5, with high-risk Mythos access still gated pending further U.S. government coordination. At launch LSVP is available through Anthropic's first-party API console and Claude for Enterprise/Team plans, but not for individual Pro/Max plans or BAA-enabled organizations. Anthropic developed the program in coordination with the U.S. government and had already onboarded dozens of organizations before opening broader applications.
Key features
First program giving vetted external access to Claude Mythos with adjusted biology safeguards
Two access tiers: Standard Use (team-wide, annual renewal) and High-risk Use (single named project, 6-month renewal)
High-risk Use currently covers Opus 5 and Sonnet 5; high-risk Mythos access remains gated pending U.S. government coordination
Developed in coordination with the U.S. government
Available via first-party console and Claude for Enterprise/Team; not yet on Pro/Max plans or for BAA-enabled organizations
Non-life-sciences safety restrictions remain in place under relaxed access
Use cases
Drug discovery research requiring less restricted biology outputs
Clinical development workflows
Biomanufacturing process design
Vetted academic and biodefense-adjacent life-sciences research
Pricing: unknown (application- and vetting-based access; no public pricing disclosed).
Microsoft merged Copilot and Microsoft 365 Copilot into one unified app, rolling out to Windows and macOS from mid-September 2026.
Details and sources
Microsoft began rolling out a unified Copilot app merging the consumer Copilot and Microsoft 365 Copilot experiences across personal, work, and school accounts, with Windows/macOS rollout starting mid-September 2026 (Windows build 19.2609.37021 released September 8, 2026). Group Chat, Podcasts, and consumer Deep Research were retired August 18, 2026, with remaining data migrated automatically.
Key features
Single unified app replacing separate Copilot and M365 Copilot apps
Automatic data migration for chats/history
New branding/logo (since August 2026)
Windows build 19.2609.37021 (Sept 8, 2026)
Use cases
Consolidated personal/work AI assistant experience
Enterprise and consumer Copilot access from one app
Pricing: unknown -- no new pricing tier disclosed; existing Copilot/M365 Copilot plans continue.
Xiaomi's open-weight omnimodal flagship, the top-scoring open-weight model on Artificial Analysis at launch.
Details and sources
Xiaomi released and open-sourced the MiMo-V2.6 series on September 22, 2026: MiMo-V2.6-Pro (its most capable model to date) and MiMo-V2.6-Flash, plus a MiMo-V2.6-Pro-UltraSpeed variant delivering up to 20x faster output at similar quality. The flagship MiMo-V2.6-Pro-RL checkpoint is a sparse Mixture-of-Experts model with 1.02 trillion total parameters (42B active per token), a 1M-token context length, and text/image/video/audio input with text output. Artificial Analysis scored it 46 on its Intelligence Index, the highest open-weight result on that leaderboard at the time. Xiaomi says it trained the model in under six days for about $2.62 million; Pro and Flash weights are released under the MIT license.
Key features
1.02T total parameters, 42B active per token (sparse MoE)
1M-token context window
Omnimodal input (text, image, video, audio), text output
MIT license for Pro and Flash weights
UltraSpeed variant: up to 20x faster output at similar quality
Evaluation, observability, and MLOps toolinginfrastructure
Mistral AI
Mistral AI Studio
Mistral's production AI platform for observability, agent orchestration, and governed prompt/skill management across enterprise deployments.
Details and sources
Mistral AI Studio launched as a private beta on October 24, 2025 as a production platform for taking enterprise AI applications from prototype to governed, monitored systems, offering an Explorer for traffic inspection, a Judge Playground for evaluation logic, an Agent Runtime built on Temporal for fault-tolerant execution, and an AI Registry for versioning agents, models, datasets, and workflows. On July 9, 2026, Mistral added version control for prompts and skills, turning them into governed production assets with immutable versions, rollback, ownership, and audit logs, so a change promoted through Studio's labels can trigger CI/CD (e.g., via SDK in GitHub Actions) and stay tied to the same asset observed in production. Studio underpins Mistral's Libraries, Vibe, and Agentic Search products for enterprise, on-premise, or self-hosted deployment.
Key features
Observability: traffic Explorer, Judge Playground, Campaigns/Datasets, and Dashboards
Agent Runtime built on Temporal for fault-tolerant, auditable agent execution
AI Registry: versioning and access control for agents, models, datasets, and judges
Version-controlled Prompts and Skills with rollback, ownership, and audit logs
Deployable in cloud, on-premise, or self-hosted/VPC environments
Use cases
Governing and versioning production prompts and agent skills
Observability and evaluation of production LLM/agent traffic
Enterprise deployment of Mistral agents with audit trails and rollback
Pricing: Enterprise platform; pricing not publicly published, sign-up/private beta at launch.
Mistral's document-parsing model that outputs bounding boxes, block classification, and confidence scores across 170 languages for enterprise RAG pipelines.
Details and sources
Mistral OCR 4, released June 23, 2026, is a document parsing and OCR model that extracts text alongside paragraph-level bounding boxes, typed-block classification (titles, tables, equations, signatures), and inline confidence scores. It ingests PDF, DOC, PPT, and OpenDocument files directly, supports 170 languages across 10 language groups, runs as a single self-hostable container, and processes up to 2,000 pages per minute on a single GPU as an ingestion layer for enterprise search, RAG, and retrieval pipelines.
Key features
Bounding boxes, typed-block classification, and per-word confidence scores
170 languages across 10 language groups
Native ingestion of PDF, DOC, PPT, and OpenDocument formats
Single-container, self-hostable deployment
Up to 2,000 pages/minute on a single GPU; 72% average human-preference win rate, OlmOCRBench score 85.20
Use cases
Document ingestion for enterprise RAG and search pipelines
Structured extraction from scanned or complex multi-format documents
Self-hosted OCR for regulated or on-premises environments
Pricing: API: $4 per 1,000 pages ($2 per 1,000 via Batch API); Document AI: $5 per 1,000 pages.
Infrastructure and inference/servinginfrastructure
Modal
Modal Shared API
Modal's OpenAI-compatible, token-billed inference endpoint, launched with day-zero Kimi K3 support.
Details and sources
Modal launched a Shared API on July 29, 2026 -- its first inference product billed by token rather than by GPU-second, a departure from its per-second compute billing model. It shipped with day-zero support (with vLLM and Moonshot) for Kimi K3, a 2.8T-parameter multimodal model with a 1M-token context window, serving it at 460 tokens/sec on release day. Kimi K3 is available via the Shared API (token pricing, Team/Enterprise) or as a dedicated Auto Endpoint on any plan.
Key features
OpenAI-compatible endpoint
Token-based billing (new for Modal)
Day-zero Kimi K3 support with vLLM/Moonshot
Dedicated Auto Endpoint option for any plan
Use cases
Serving open multimodal models without managing GPU infrastructure
Cost-predictable token billing vs GPU-second billing
Fast access to newly released open-weight models
Pricing: Kimi K3: $3/M prompt tokens, $0.30/M cached prompt tokens, $15/M completion/reasoning tokens on Shared API (Team/Enterprise); Auto Endpoint available on any plan.
MCP's anniversary spec revision makes the protocol stateless, cacheable, and header-routable, with hardened OAuth and a formal async Tasks extension.
Details and sources
Published July 28, 2026, this MCP revision removes the initialize/initialized handshake and session IDs for self-describing stateless requests, adds Multi Round-Trip Requests (MRTR) for confirmations without held-open streams, and moves method/tool names into HTTP headers (Mcp-Method, Mcp-Name) for gateway/WAF routing. It hardens auth with RFC 9207 issuer validation and Client ID Metadata Documents replacing Dynamic Client Registration, adds ttlMs/cacheScope caching hints, formalizes the Tasks extension for long-running async operations, and deprecates Roots/Sampling/Logging with a 12-month support window. All Tier 1 SDKs (TypeScript, Python, Go, C#) support it at launch; Rust support is in beta.
Key features
Stateless request/response core, no session IDs or handshake
Multi Round-Trip Requests (MRTR) replace server-initiated confirmations
Header-based routing (Mcp-Method, Mcp-Name) for gateways and WAFs
RFC 9207 issuer validation plus Client ID Metadata Documents replace Dynamic Client Registration
Tasks extension formalized for long-running async operations
Use cases
Building MCP gateways/routers needing cacheable, statelessly-routable requests
Enterprises needing stricter OAuth/issuer validation for MCP servers
Anthropic's specification letting AI agents safely operate physical lab and manufacturing hardware, extending agent tool-use to the physical world.
Details and sources
Announced August 27, 2026, the Model Hardware Standard (MHS) is a shared specification that lets AI agents read sensors, write to actuators, and operate physical equipment such as microscopes, liquid handlers, robotic arms, and lasers through a common interface, cutting hardware integration time from weeks or months down to hours or minutes. Anthropic opened a research preview with named partners Genentech, the University of Washington's Baker and Pinglay labs, Carnegie Mellon University, HHMI Janelia Research Campus, QuEra Computing, and Tetsuwan Scientific. Early results include QuEra improving laser-lock recovery success from 58% to 99.3% (cutting recovery time from 150 seconds to 6 seconds) and Carnegie Mellon cutting serial-dilution experiment time roughly 3x while achieving curve-fitting accuracy above R² 0.98. Anthropic plans to open source the full specification after developing safety evaluations and best practices with launch partners.
Key features
Common interface for AI agents to read sensors and operate actuators on lab/manufacturing hardware
Cuts hardware integration time from weeks/months to hours/minutes
Enables collision-free, parallel operation of multiple instruments (demonstrated at University of Washington)
Research preview partners: Genentech, University of Washington, Carnegie Mellon, HHMI Janelia Research Campus, QuEra Computing, Tetsuwan Scientific
QuEra Computing: laser-lock recovery success up from 58% to 99.3%, recovery time cut from 150s to 6s
Full specification planned to be open sourced after safety evaluation with launch partners
Use cases
Automating drug-discovery and liquid-handling lab experiments (Genentech)
Serial-dilution experiment automation with curve-fitting validation (Carnegie Mellon)
Precision calibration tasks such as quantum-computer laser locking (QuEra Computing)
Collision-free multi-instrument coordination in research labs (University of Washington)
Pricing: unknown (research preview with a vetted partner cohort; no public pricing disclosed).
Meta's personal AI agent that completes real tasks like booking and shopping on a dedicated secure cloud VM, not just a chatbot.
Details and sources
Meta launched Muse on September 8, 2026, describing it as a secure, private personal AI agent that proactively helps with people's goals rather than just answering questions. Muse runs on a dedicated Muse Secure VM with its own browser, storing a user's connected-app credentials there, and a Sentinel sub-agent requires approval before sensitive actions like payments. It is powered by Meta's Muse Spark model, remembers user preferences to make unprompted suggestions, and connects to payment rails including Link by Stripe (with Shop Pay coming). Muse rolled out first in the US on iOS, Android, muse.ai, and WhatsApp, passed 2.5 million downloads within two weeks per Sensor Tower, and has drawn friction from Amazon, which blocked Muse from completing purchase transactions on Amazon.com. Meta shipped a native macOS app for Muse around September 17-18, 2026, giving the agent access to the Mac's Files, Mail, Messages, Calendar, and Notes apps so it can sort files, pull information out of email threads, add calendar events, and draft Messages replies on the user's behalf; Mac actions run through the same cloud Muse Secure VM, with the Sentinel gating layer still requiring explicit approval before sensitive outbound actions like file deletion, email sends, or purchases.
Key features
Runs on a dedicated Muse Secure VM with its own browser and securely stored app credentials
Sentinel sub-agent requires explicit approval before sensitive actions such as payments, purchases, file deletion, or sending email
Persistent memory that remembers preferences (e.g., dietary restrictions) and proactively suggests actions
Payment integration via Link by Stripe, with Shop Pay integration coming
Continues working asynchronously in the background after the app is closed
Powered by Meta's Muse Spark model; planned Muse Confidential VM will add end-to-end encryption of VM data with a user-held key
Native macOS app (Muse for Mac) with access to Files, Mail, Messages, Calendar, and Notes, acting on the user's behalf across those native apps
Use cases
Booking travel, movie tickets, and appointments end to end
Turning a saved recipe or Instagram post into a grocery list or dinner-party menu
Automating recurring personal admin like permission slips and scheduling
Agentic online shopping and checkout across connected merchants
Desktop file organization, email triage, and calendar/message management on Mac
Pricing: Free tier (100 million Muse tokens/week); Power tier $20/month (500 million tokens/week); Maximum tier $100/month (3 billion tokens/week).
Meta's open-weight 30B-parameter agentic model, distilled from Muse Spark, that runs locally on consumer hardware with tool use and multimodal input.
Details and sources
Meta released Muse Glimmer on Hugging Face on August 10, 2026 (huggingface.co/blog/muse-glimmer), an open-weight 30-billion-parameter causal language model (a 2B vision encoder plus a 28B text decoder) distilled from Muse Spark and purpose-built for autonomous agentic tasks on consumer hardware, under an Apache 2.0 license. It combines multi-step reasoning, tool use, multimodal understanding (via a dedicated perception encoder for interleaved text and images), and failure recovery in a single model that runs locally without cloud infrastructure, using a hybrid sliding-window/full-attention architecture, grouped-query attention, and DFlash speculative decoding for faster generation. Meta reported scores of 75.5 on MCP-Atlas, 74.6 on DeepSearch QA, and 51.2 on SWE-Bench Pro, and shipped day-0 support across Transformers, llama.cpp (including GGUF quantized builds), vLLM, and Inference Endpoints, plus training data spanning more than 100 languages.
Meta Superintelligence Labs' first image-generation model, which invokes search and coding tools to self-refine outputs rather than mapping prompts directly to pixels.
Details and sources
Meta introduced Muse Image on July 7, 2026, its first image-generation model built by Meta Superintelligence Labs (internally codenamed "Mango"). Instead of a direct prompt-to-image mapping, Muse Image operates agentically: it invokes search and coding tools to improve accuracy, self-refines its own generations, and improves via scaling test-time compute; it supports multi-reference composition and precise iterative editing, and integrates with Muse Spark for joint planning. Meta reported it ranked No. 2 on the Arena leaderboard for text-to-image, single-image editing, and multi-image editing as of July 5, 2026, and applies invisible Content Seal watermarking to outputs. It shipped in the Meta AI app, meta.ai, Instagram Stories (US), and WhatsApp (limited countries), with Facebook support planned, and is priced on the Meta Model API at $0.01 per image.
Key features
Agentic generation: invokes coding and web-search tools to improve accuracy and self-refine images
Multi-reference image composition and iterative, precise editing
Integrates with Muse Spark for joint text/image planning
Invisible Content Seal watermarking on generated images
Ranked No. 2 on Arena for text-to-image and image-editing categories (as of July 5, 2026)
Available on Meta Model API at $0.01/image
Use cases
Generating and editing images inside Meta AI, Instagram, and WhatsApp
Composing new images from multiple reference photos
Developer access to image generation via the Meta Model API
Pricing: $0.01 per image via the Meta Model API; free within Meta AI app, meta.ai, Instagram Stories (US), and WhatsApp (limited countries).
Meta's hosted frontier reasoning model with a 1M-token context window, built for agentic and computer-use workflows, launched with the Meta Model API.
Details and sources
Meta released Muse Spark 1.1 on July 9, 2026, a multimodal reasoning model built for agentic tasks with a 1-million-token context window with active context management. Unlike Llama, which Meta positions as open-weight models for local use and fine-tuning, Muse Spark is a hosted, high-performance model aimed at computer-use workflows that span multiple applications, adapt to evolving requirements, and navigate unfamiliar interfaces; independent benchmark coverage put its reasoning, coding, and instruction-following scores on par with GPT-5.5. Meta paired the release with a public preview of the Meta Model API (OpenAI-compatible interface, $1.25/million input and $4.25/million output tokens, $20 in free credits) and made the model available in "Thinking" mode in the Meta AI app and on meta.ai. Muse Spark 1.1 also underpins Meta's Muse personal agent and Muse Image; Meta has since shipped a coding-focused Muse Spark 1.2 (with the companion Muse Code, announced August 5, 2026) and a subsequent Muse Spark 1.3.
Key features
1,000,000-token context window with active context management
Strong tool use, computer-use, coding, and multimodal understanding per Meta's release benchmarks
Zero-shot generalization to new tools and custom skills; multi-agent orchestration support
Meta Model API public preview with OpenAI-compatible interface
Powers Meta's Muse personal agent and integrates with Muse Image for joint planning
Superseded within weeks by coding-focused Muse Spark 1.2 (with Muse Code) and Muse Spark 1.3
Long-context document and multi-step reasoning tasks
Developer access to a frontier-class hosted model via API
Backbone model for Meta's Muse and Muse Image products
Pricing: Meta Model API: $1.25/million input tokens, $4.25/million output tokens, with $20 in free credits; also available free in "Thinking" mode in the Meta AI app and on meta.ai.
Meta Superintelligence Labs' text-to-video model with native audio, built on the same pretraining base as Muse Image and previewed alongside it.
This entry is marked for review.
Details and sources
Meta shared an early preview of Muse Video on July 7, 2026, alongside Muse Image, as the second of Meta Superintelligence Labs' first media-generation models. Built on the same pretraining base as Muse Image, Muse Video generates video with native audio support and, per Meta, ranked No. 3 in human-preference Elo on the Arena text-to-video leaderboard at the time of release. Meta said it is investing further in areas with current performance gaps, including audio-video synchronization and physically accurate fast motion, and plans to extend Content Seal invisible watermarking to video. As of the preview, Muse Video was not yet broadly released, with Meta describing it as coming soon to creators and Meta AI.
Key features
Native audio generation alongside video output
Shares pretraining base with Muse Image
Ranked No. 3 in human-preference Elo on Arena's text-to-video leaderboard at preview
Content Seal invisible watermarking planned for video output
Early preview only; broader rollout to creators and Meta AI still pending as of the announcement
Use cases
Text-to-video generation with synchronized audio for creators
Planned integration into the Meta AI app for video content creation
Pricing: Not yet disclosed; announced as an early preview, coming soon to creators and Meta AI (not broadly available at launch).
NVIDIA's open-weight model, downloadable and modifiable without a license fee.
This entry is marked for review.
Details and sources
NVIDIA released Nemotron 3.5 Lightning on August 11, 2026, an open-weight AI model that anyone can download, run, and modify without paying a license fee, continuing NVIDIA's Nemotron 3 family strategy of releasing weights, training data, RL environments, and post-training recipes.
Key features
No license fee for download, run, or modify
Part of the broader Nemotron 3 open family (alongside Nano, Super, Ultra)
NVIDIA also releases training data and RL environments alongside weights
Use cases
Enterprise on-prem deployment
Research requiring full training-recipe transparency
Pricing: Free, open weights (exact license name unconfirmed).
Cohere Labs' 2.4B-parameter open-weight vision-language model built for native-resolution document understanding on edge hardware.
Details and sources
North Micro Vision Instruct is a 2.4-billion-parameter open-weight vision-language model released by Cohere Labs on August 12, 2026 under Apache 2.0. It combines a native-resolution vision encoder, a projector, and a compact language model to process images up to 1654x2339 pixels (A4 documents at 200 dpi) without downscaling, targeting document understanding, visual grounding, OCR, and multilingual chart/document reasoning small enough for local and edge deployment.
Key features
2.4B parameters, Apache 2.0 licensed weights
Native-resolution vision encoder supports up to 1654x2339 px (A4 @ 200 dpi)
Strong document understanding, visual grounding, and OCR performance
Multilingual visual understanding across documents, charts, and natural images
Fine-tuning support via NVIDIA AutoModel and Axolotl; MLX-VLM quantizations available
Use cases
On-device or edge document OCR and understanding
Multilingual chart and document question answering
Fine-tuning a compact VLM for domain-specific visual tasks
Pricing: Free and open source (Apache 2.0); weights hosted on Hugging Face.
Notion's Sept 2026 release adds shareable AI Skills, sub-agent delegation, and a public Agent SDK.
Details and sources
Notion 3.7, released September 15, 2026, introduces a Skills library of reusable AI instructions that can be downloaded into Claude Code, Codex, Cursor, Gemini, or Grok. It adds sub-agent delegation (Custom Agents can spawn other Custom Agents with distinct instructions/models), custom MCP connections to GitHub and Amplitude, an iOS Notion Agents app, and a public-beta Agent SDK plus Developer Portal for managing Workers/connections. AI Search is also 50% faster.
Key features
Skills library -- reusable AI instructions shareable across Claude Code, Codex, Cursor, Gemini, Grok
Sub-agents: Custom Agents can delegate to other Custom Agents with distinct context/models
Custom MCP connections to GitHub and Amplitude (Business/Enterprise beta)
Agent SDK in public beta; new Developer Portal in the Notion sidebar
AI Search 50% faster; model picker comparing Opus 5, GPT-5.6 Sol, Kimi K3
Use cases
Standardizing how a team's AI agents perform recurring tasks
IT help-desk automation via delegated sub-agents
Post-meeting recap/action-item automation
Embedding Notion workflows into external coding agents via Skills
Pricing: Most features on Business/Enterprise plans; Agent SDK and Developer Portal in public beta, no separate pricing disclosed.
Ollama v0.34.0 brings official ChatGPT Desktop integration on macOS, letting local models run inside ChatGPT's app.
Details and sources
Released September 5, 2026, Ollama v0.34.0 adds official ChatGPT Desktop integration (macOS) so local Ollama models run directly inside the ChatGPT desktop app, improves structured-output performance on Apple Silicon, and adds OpenAI-compatible client tool search and response compaction.
Key features
ChatGPT Desktop integration (macOS)
Faster structured output on Apple Silicon
OpenAI-compatible client tool search
Response compaction
Use cases
Running local/open-weight models inside the ChatGPT Desktop workflow
OpenAI's managed enterprise platform for deploying governed voice and chat agents into high-volume, high-stakes production workflows.
Details and sources
OpenAI Presence, announced July 22, 2026, is a managed platform for building, deploying, operating, and continuously improving governed AI agents for high-volume, high-stakes enterprise workflows such as billing issues, insurance claims, or IT service requests. Each deployment is scoped to a specific job, giving the agent only the knowledge and system access that job requires, while the enterprise sets policies for what the agent can do, when it needs approval, and when to escalate to a human. Presence is offered through a limited general-availability program, with deployments led by OpenAI Forward Deployed Engineers and select systems integrators; it is not yet a self-serve product.
Key features
Job-scoped agents with least-privilege knowledge and system access
Built-in guardrails, policies, and escalation rules with accuracy/performance verification
Managed by OpenAI Forward Deployed Engineers and select global systems integrators
Designed for high-volume, high-stakes voice and chat workflows
Use cases
Contact-center-style billing and insurance claim resolution
Employee IT service request handling
Enterprise voice/chat agents that must escalate to humans under defined policy
Pricing: Limited general-availability program for eligible enterprise customers; not self-serve, pricing not publicly listed.
Opera's agentic AI browser dropped its paid-only model to become a free download with an MCP connector and CLI.
Details and sources
In August 2026, Opera Neon changed its business model from paid-only to a free download, letting users plug existing AI agents (Claude, ChatGPT, etc.) into the browser via its MCP connector and a new Opera Browser CLI, rather than requiring Opera's own bundled agent credits.
Key features
Free tier (previously paid-only)
MCP connector: acts as an MCP server for external AI clients
New Opera Browser CLI
Auto-selects among built-in agents (Chat, Do, Make, 1 Minute Research)
Use cases
Bring-your-own-agent agentic browsing
Automating in-browser tasks via external AI clients
Pricing: Free tier available since August 2026; a paid Standard plan (around $19.90, per third-party review) also exists for expanded usage.
Perplexity's agentic AI browser, now on all major platforms, with local-desktop context via Personal Computer for Mac.
Details and sources
Perplexity's Comet browser reached full platform parity in 2026 and gained several agentic-computer features in this window: Personal Computer for Mac (connects local files, apps, and Comet context to Perplexity Computer) and full iOS availability in August 2026, and, controversially, moved the 'Control browser' capability behind separate Computer credits around September 15, 2026 after it had shipped free with Pro.
Key features
Personal Computer for Mac: local file/app context for Computer
Comet available on iOS, Android, Mac, and Windows
Computer effort controls (Light to Ultra) for model/reasoning selection
Pricing: Comet free tier exists; advanced 'Computer' agent actions (including 'Control browser') now consume separate Computer credits as of Sept 2026, a change from being bundled free with Pro.
Pinecone's knowledge-orchestration layer that compiles enterprise data into cited artifacts agents can query directly.
Details and sources
Pinecone Nexus reached general availability in August 2026, repositioning Pinecone from a vector search index into an agent knowledge engine. It compiles scattered enterprise data sources into cited, queryable artifacts rather than raw vector matches, and ships alongside a new Marketplace of prebuilt AI application templates and a $20/month flat Builder pricing tier.
Key features
Compiles enterprise data into cited artifacts for agent queries
Marketplace of prebuilt AI application templates
Builder tier: $20/month flat, 10 serverless indexes, 10GB storage, 100 namespaces per index
Cohere Rerank 4.0 Fast integration
Serverless expansion to AWS Frankfurt and Singapore
Built-in Inference API for embedding generation
Use cases
Agent knowledge retrieval with citations
RAG at enterprise scale
Cost-tiered vector search for small teams
Pricing: Builder tier $20/month flat; Starter free with hard limits; Standard usage-based.
A GitHub Copilot CLI research preview that routes each coding task across multiple AI models for frontier quality at lower cost.
Details and sources
Announced September 4, 2026, HydraFusion builds a runtime execution plan per task in Copilot CLI, choosing among three workflow patterns -- Single, Cascade with a quality gate, and Critique with a cross-family reviewer -- across models from multiple providers. In GitHub's offline evaluations across three agentic coding benchmarks it improved verified task quality by 4.9 percentage points at an estimated 67% lower cost than Claude Opus 5 on TerminalBench 2.1. It is selectable via /experimental in Copilot CLI on all plans, billed at each underlying model's standard rate.
Key features
Three workflow patterns: Single, Cascade with a quality gate, Critique with a cross-family reviewer
Runtime-built execution plan per task rather than a fixed router
+4.9 percentage points verified quality, ~67% lower cost vs Claude Opus 5 on TerminalBench 2.1 (GitHub's own evals)
Selectable via /experimental in Copilot CLI
Use cases
Agentic coding needing frontier quality at controlled cost
Teams wanting automatic per-task model selection
Pricing: No separate fee; billed at the underlying model's standard token rate, available on all Copilot plans.
Alibaba's flagship 2.4T-parameter model, the first open-weight Max-tier Qwen release.
This entry is marked for review.
Details and sources
Qwen3.8-Max released August 3, 2026 at 2.4 trillion parameters (95B active), priced at $2/million input and $6/million output via API. Alibaba published its weights on August 12, 2026 as a text-only variant under a bespoke Qwen3.8-Max license (not Apache 2.0) -- the first open Max-tier Qwen release.
Key features
2.4T parameters, 95B active (Mixture-of-Experts)
First open-weight release of a Max-tier Qwen model
Text-only open variant under a bespoke license (not Apache 2.0)
Use cases
High-end reasoning and agentic tasks via API
Self-hosted deployment of open Max-tier weights (text-only)
Alibaba's omnimodal model taking text, image, audio, and video in one request with a 1M-token context window.
Details and sources
Qwen3.8-Omni-Flash is Qwen's native omnimodal model, released September 18, 2026, unifying text, image, audio, and video understanding in a single model with a 1-million-token context window. It targets agentic audio-video understanding and tool use at lower cost than Qwen's larger omni models.
Key features
1M-token context window
Native text, image, audio, and video input in one request
Replit added automatic model routing, enterprise governance, pen-testing, and GA Databricks Lakebase support across August-September 2026.
This entry is marked for review.
Details and sources
Across August-September 2026 Replit shipped: Intelligent Model Routing that auto-selects the best model and effort level per task; enterprise governance controls; on-demand black-box penetration testing simulating real attacks on deployed apps; and, September 10, general availability of the Replit-Databricks integration with native Lakebase support, auto-provisioning a Lakebase database on deploy.
Key features
Intelligent Model Routing picks model and effort level automatically
Enterprise governance controls for scaled teams
Black-box penetration testing on deployed apps
GA Databricks integration, auto-provisioned Lakebase database on deploy
Use cases
Non-developers building full-stack apps from natural language
Enterprise teams needing governance and security testing
Apps needing a governed production Postgres-compatible database provisioned automatically
Pricing: unknown for these specific features; layered into existing Core/Pro/Enterprise tiers.
Mistral's 8B robotics navigation model that steers wheeled, legged, or flying robots using a single RGB camera and plain-language instructions.
Details and sources
Robostral Navigate, released July 8, 2026, is an 8-billion-parameter embodied-AI model that lets robots autonomously navigate complex environments from a single RGB camera and a natural-language instruction, with no LiDAR or depth sensors required. Trained entirely in simulation on roughly 2.4 million recorded trajectories across thousands of virtual spaces, it achieves 76.6% success on unseen R2R-CE navigation benchmarks, outperforming both the best single-camera approach and systems using depth or multiple cameras, and generalizes across wheeled, legged, and flying robot form factors. It is a proprietary, enterprise-access model marking Mistral's entry into physical AI for factories, warehouses, and industrial automation; weights are not publicly released.
Key features
Single RGB camera input, no depth sensors or LiDAR required
76.6% success on R2R-CE validation unseen; beats best single-camera approach by 9.7 points and depth/multi-camera systems by 4.5 points
Generalizes across wheeled, legged, and flying robot hardware
Trained entirely in simulation (~2.4M trajectories across 6,000+ virtual spaces)
Combines pointing-based navigation with displacement commands and online reinforcement learning
Use cases
Autonomous robot navigation in offices, warehouses, and industrial facilities
Vision-language navigation for indoor/outdoor service robots
Physical AI deployments across mixed robot fleets and form factors
Pricing: Not publicly priced; enterprise access via contact/sales, no open weights or public API at launch.
A CrowdStrike-NVIDIA agentic system built on Nemotron models for autonomous cyber defense, including AI-vs-AI red/blue teaming.
This entry is marked for review.
Details and sources
Also unveiled at Fal.Con 2026 (September 2026), SafeMind is described as the first agentic system purpose-built for cybersecurity defenders, jointly developed by CrowdStrike and NVIDIA using NVIDIA Nemotron models trained for cybersecurity, combined with CrowdStrike's security telemetry. It runs adversarial red-team/blue-team simulations via digital twins and aims to detect and respond to threats autonomously at machine speed.
Key features
Built on NVIDIA Nemotron frontier models tuned for cybersecurity
Adversarial red-team/blue-team simulation via digital twins
Autonomous, machine-speed detect-and-respond loop
Joint CrowdStrike/NVIDIA data and model integration
Use cases
Continuous AI-vs-AI defensive testing for SOC teams
Automated threat response without human-speed bottlenecks
ElevenLabs' batch speech-recognition model specialized for medical and clinical audio, now generally available.
Details and sources
ElevenLabs made Scribe v2 Medical generally available on September 11, 2026. It is a batch speech-to-text model tuned for medical and clinical audio transcription, billed at the same rate as the standard Scribe v2 model.
Key features
Specialized for medical/clinical audio
Batch (non-realtime) speech recognition
Same pricing tier as Scribe v2
Use cases
Clinical documentation transcription
Medical dictation workflows
Pricing: Same rate as Scribe v2 (exact per-minute rate unconfirmed).
Mistral's 3B open-weight, policy-adaptive safety classifier for text and image moderation.
Details and sources
Shieldstral 1.0, released August 4, 2026, is a 3-billion-parameter open-weight multimodal safety classifier that judges text and images against moderation policies written in plain language at inference time, rather than fixed categories. Mistral says it matches or beats open guard models up to 7x its size on text safety and sets a new state of the art on multimodal moderation, running on a single 16GB GPU.
Key features
3B parameters, runs on one 16GB GPU
Policy-adaptive: takes a moderation policy as plain-language input
Multimodal (text + image) safety scoring
Covers 12 languages
Use cases
On-device content moderation
Agent guardrails with a custom safety policy
Pricing: Apache 2.0, weights free on Hugging Face.
Slack workspace feature turning channels into multiplayer environments for AI coding agents like Claude Code, Devin, and Copilot.
Details and sources
Announced August 24, 2026, Slack Code auto-provisions project-based code channels when an AI coding agent is tagged, giving a shared space with conversation, an agent action-plan tab, line-by-line diffs, and a live output preview. It supports Anthropic's Claude Code, Cognition's Devin, GitHub Copilot, OpenAI's ChatGPT, and Vercel agents, with human-in-the-loop pause/redirect controls, enterprise governance, and searchable audit logs.
Key features
Auto-provisioned code channels on agent tag
Four-tab view: conversation, action plan, diffs, live preview
Human-in-the-loop pause/redirect of agents mid-task
Enterprise governance and searchable audit logs
Supports Claude Code, Devin, GitHub Copilot, ChatGPT, and Vercel agents
Use cases
Cross-functional teams reviewing AI-agent coding work live in Slack
Governed, auditable agentic development for regulated organizations
Bringing PMs/designers into a coding agent's workflow without a separate IDE
Pricing: Included across all Slack plans; each partner coding agent requires its own separate access/license.
Zendesk's industry- and business-specific AI agents, launching in commerce, that automate up to 80% of workflows.
Details and sources
Zendesk introduced Specialized AI Agents as purpose-built agents that combine industry expertise with a company's own knowledge, workflows, and connected systems to take action, which Zendesk says can automate up to 80% of relevant workflows. The launch includes two types: Industry Agents, pre-built for common high-value work in a given industry, and Custom Agents, which businesses configure around their own processes and policies. Zendesk is launching first in commerce, where the agents help customers shop, manage orders, process returns and exchanges, resolve delivery issues, and handle refunds, with financial services, media, and technology industries planned next. The launch builds on Zendesk's Autonomous Service Workforce vision first introduced at Relate 2026.
Key features
Industry Agents: pre-built for common high-value work in a specific industry
Custom Agents: configured by businesses around their own processes and policies
Launching first in commerce (shopping, order management, returns/exchanges, refunds, delivery issues)
Claims automation of up to 80% of relevant workflows
Part of Zendesk's broader Autonomous Service Workforce vision
OpenAI's in-ChatGPT ad format that opens a labeled, direct conversation with a brand's own AI agent instead of linking out.
Details and sources
OpenAI formalized Sponsored Agents on September 16, 2026, an advertising format that lets a ChatGPT user click an ad and start a direct, labeled conversation with an advertiser's AI agent inside the ChatGPT interface, rather than following an external link. Users can ask personalized questions about the advertised product and get tailored recommendations and links back to the brand's site. It is currently being tested with select advertisers in the United States. Alongside Sponsored Agents, OpenAI introduced an Ads Manager plugin for natural-language campaign creation and analysis, plus its first CRM and ecommerce integrations, HubSpot and Shopify. OpenAI's advertising division reportedly reached a $1 billion annualized revenue run rate in under 200 days.
Key features
In-conversation, labeled sponsored agent chats instead of outbound ad links
Ads Manager plugin for natural-language campaign creation and analysis
First CRM (HubSpot) and ecommerce (Shopify) advertiser integrations
Currently limited to select US advertisers in testing
Use cases
Brands offering an interactive product Q&A agent directly inside a ChatGPT ad
Advertisers managing and analyzing ChatGPT ad campaigns via natural language
CRM/ecommerce-connected ad campaigns through HubSpot or Shopify
Pricing: Advertiser-funded ad format currently in limited testing with select US advertisers; pricing/rate card not publicly listed.
Suno's first fully-licensed AI music model family, built with Warner Music, BMG, Believe and TuneCore.
Details and sources
Suno launched v6 on September 9, 2026, its first music-generation model developed with major rights holders after settling prior copyright disputes. The family ships as three tiers: v6 and v6-wild for Pro/Premier subscribers, and v6-mini across all plans, all faster and higher-quality than prior models, which have been fully retired for new generations.
Key features
Three-tier model family: v6, v6-wild, v6-mini
First model built with licensed inputs from Warner Music, BMG, Believe, and TuneCore
All pre-v6 models retired for new generation work
Greater creator control over structure and expression
Use cases
Licensed commercial music generation
Songwriting and production assistance
Label-backed AI music releases
Pricing: v6/v6-wild require Pro or Premier subscription; v6-mini available on all plans (exact prices not published).
A desktop workspace, from the team behind Module Federation, where teams and reusable AI 'Specialists' collaborate in shared channels and build workflow mini-apps.
This entry is marked for review.
Details and sources
Zephyr reached general availability of The AI Platform (TAP) on August 20, 2026, after roughly five months in early access, describing it as a desktop app where entire teams collaborate with AI 'Specialists' (reusable agents with durable roles, instructions, tools, and knowledge) in shared, Slack-style channels rather than one-off chatbot windows. Its centerpiece is the Mini-App Builder, which turns a repeatable workflow into an AI-enabled application the whole company can install; Zephyr runs its own CRM, content pipeline, and scheduling as mini-apps on TAP as a demonstration. A provider-agnostic Model Router lets teams avoid vendor lock-in by routing across OpenRouter, OpenAI, Anthropic Claude, Google Gemini, AWS Bedrock, and local or self-hosted endpoints. TAP is available for macOS, Windows, and Linux at theaiplatform.app. Zephyr was founded by the team behind the Module Federation project.
Key features
Slack-style channels where humans and reusable AI 'Specialist' agents work side by side
Mini-App Builder: turns a repeatable workflow into an installable AI-enabled app
Provider-agnostic Model Router across OpenRouter, OpenAI, Anthropic, Google Gemini, AWS Bedrock, and self-hosted endpoints
Desktop app for macOS, Windows, and Linux
Reached GA after ~5 months in early access; founded by the Module Federation team
Use cases
Team-wide collaboration with persistent, role-based AI agents instead of ad hoc chat sessions
Turning recurring workflows (sales, content, scheduling) into internal mini-apps
Avoiding vendor lock-in by routing tasks across multiple model providers
Pricing: Not publicly disclosed in available sources at GA launch.
Thomson Reuters' first in-house proprietary large language model, trained on its Westlaw and Practical Law content for professional-grade accuracy.
Details and sources
Thomson Reuters announced Thomson, its first proprietary large language model built in-house, on August 24, 2026. The company invested $40 million training it on an open-source foundation combined with decades of proprietary content from Westlaw and Practical Law, targeting what it calls "Fiduciary-Grade" accuracy for professional and legal work at lower inference cost than typical frontier models. Thomson Reuters says the model was trained on less than 10% of its content library so far and is deployed within CoCounsel Legal as part of a multi-model approach, with an open-weight release on Hugging Face for academic, non-commercial use.
Key features
Trained on proprietary Westlaw and Practical Law content plus an open-source base
Built to Thomson Reuters' "Fiduciary-Grade" accuracy standard for professional work
Deployed inside CoCounsel Legal alongside other models in a multi-model strategy
Open-weight release on Hugging Face for academic, non-commercial use
Designed for sovereign AI: full control over training data, behavior, and privacy
Use cases
Legal research and drafting inside CoCounsel Legal
Domain-specific professional reasoning tasks requiring high factual accuracy
Academic research use via the open-weight release
Pricing: Not publicly disclosed for the commercial/CoCounsel deployment; open-weight version available free on Hugging Face for academic, non-commercial use.
Together AI Fine-Tuning Platform (Expert LoRA / DPO update)
Together AI's hosted fine-tuning platform adds a no-code UI, Direct Preference Optimization, continued training, and Expert LoRA for MoE models.
Details and sources
On September 11, 2026, Together AI shipped a major update to its hosted fine-tuning platform, adding a no-code web UI (previously SDK/API-only), Direct Preference Optimization and continued-training support, and 'Expert LoRA' -- a Mixture-of-Experts-specific low-rank adaptation technique for fine-tuning sparse MoE models like GLM-5.3. New model support includes Gemma 3 and distilled DeepSeek-R1 variants.
Transformers v5.17.0 adds seven new model architectures, including Tencent's 780B HYV4 MoE and Moonshot's Kimi Linear attention.
Details and sources
Released September 9, 2026, Hugging Face's Transformers v5.17.0 adds seven new architectures -- HYV4 (Tencent, 780B-parameter MoE), VibeVoice, NeoMME, Fun-ASR-Nano, Kimi Linear (Moonshot), Canary-1B-v2, and NeuCodec -- plus generation, cache, kernel, quantization, and vision-RoPE improvements, with accompanying breaking changes.
Key features
7 new architectures added (HYV4 780B MoE, VibeVoice, NeoMME, Fun-ASR-Nano, Kimi Linear, Canary-1B-v2, NeuCodec)
Generation/cache/kernel/quantization improvements
Vision RoPE improvements
Documented breaking changes
Use cases
Running/fine-tuning newly supported open-weight architectures via transformers
Infrastructure and inference/servinginfrastructure
vLLM Project
vLLM
Open-source LLM inference/serving engine; shipped v0.28.0 through v0.30.0 in this window, adding a persistent-GPU-daemon cold-start feature and DeepSeek-V4.1-Flash support.
Details and sources
vLLM shipped v0.28.0 (August 26, 2026), v0.29.0 (September 9, 2026), and v0.30.0 (September 22, 2026, 762 commits from 315 contributors) in this window. v0.30.0 adds 'Fast Start,' a persistent per-GPU daemon that holds pre-quantized, tensor-parallel-sharded weights in GPU memory and maps them via CUDA IPC instead of disk reads; Gumbel-max watermarking to mark and detect vLLM-generated text; CUDA graph capture time cut from 12s to 2s and engine init from 28.9s to 8.2s on H200; and DeepSeek-V4.1-Flash support with FlashML MXFP8 KV caching. It carries several breaking changes: scale-out endpoints are now opt-in, GPTQ g_idx support was removed, and the grpc entrypoint moved to `vllm serve --grpc`.
Key features
Fast Start: persistent GPU daemon with CUDA IPC weight mapping for near-instant cold starts
Gumbel-max watermarking for AI-generated text detection
CUDA graph capture 12s to 2s, engine init 28.9s to 8.2s on H200
DeepSeek-V4.1-Flash support via FlashML MXFP8 KV cache
Regular release cadence with DeepSeek-V4 optimizations and tiered KV offloading
Use cases
Self-hosted high-throughput LLM serving
Fast-cold-start autoscaling deployments
Open-weight model deployment at scale
Detecting AI-generated text from self-hosted models
Mistral's 24B open-weight speech-to-text model for transcription, translation, and audio understanding.
This entry is marked for review.
Details and sources
Voxtral Small 24B 2507 STT is Mistral AI's open-weight speech transcription model suited for transcription, translation, and audio understanding workloads.
Key features
24B parameters
Speech-to-text, translation, and audio understanding
VS Code 1.138 lets AI coding agents run inside a project's local Dev Container, keeping agent work aligned to the project's own toolchain and dependencies.
Details and sources
Visual Studio Code 1.138, released mid-September 2026, adds the ability to run agent sessions inside a local Dev Container, so an agent uses the environment and dependencies configured for the project rather than everything installed on the local machine (Docker is required). The release also expands the Codex harness so sessions can continue across apps and users can switch between Copilot and ChatGPT subscriptions from the model picker without losing the conversation, and adds a preview session-cleanup feature that marks merged sessions done and can delete them after a grace period.
Key features
Agent sessions run inside a project's local Dev Container (requires Docker)
Gated behind the chat.agentHost.devContainer setting, which adds a "Use Dev Container" folder action
Expanded Codex harness: sessions continue across apps
Switch between GitHub Copilot and ChatGPT subscriptions from the model picker mid-conversation
Preview session cleanup: auto-marks merged sessions done, optional deletion after a grace period
Use cases
Keeping AI agent coding work isolated to a project's configured toolchain and dependencies
Continuing an agent coding session across different apps/devices
Managing and cleaning up completed agent sessions in the Agents window
Pricing: Included with VS Code (free); underlying Codex/Copilot/ChatGPT usage billed per the user's existing subscription.
Infrastructure and inference/servinginfrastructure
Cloudflare
Workers AI Model Catalog Additions (Aug 2026)
Cloudflare added six new open-weight text models to Workers AI, including DeepSeek-V4 and Kimi K2.7-code.
Details and sources
On August 26, 2026, Cloudflare added six new text-generation models to Workers AI: DeepSeek-V4-Flash and DeepSeek-V4-Pro (1,048,576-token context), gpt-oss-120b and gpt-oss-20b (128,000-token context), Qwen3.8-27B, and Kimi-K2.7-code (262,144-token context each). All run natively on Workers AI with no external provider key required.
Key features
DeepSeek-V4-Flash/Pro with 1M-token context
gpt-oss-120b/20b
Qwen3.8-27B, vision-capable
Kimi-K2.7-code for coding tasks
No external API key needed -- billed through Workers AI
Use cases
Edge inference for long-context agentic workflows
Cost-efficient access to open-weight frontier-adjacent models
Coding assistants via Kimi-K2.7-code
Pricing: Billed through Workers AI standard pricing; no additional provider key or cost required per model.
Huawei's next-generation AI accelerator, its commercial launch moved up to Q1 2027 as the anchor chip of the Atlas 950 SuperPoD/SuperCluster.
Details and sources
Huawei announced at Huawei Connect 2026 in Shanghai on September 17, 2026 that it is accelerating commercial availability of its Ascend 960DT AI accelerator from late 2027 to Q1 2027, citing AI chip demand outstripping its manufacturing capacity. Rotating chairman David Wang unveiled the chip alongside a companion variant, the Ascend 960PR, as the basis for Huawei's Atlas 950 SuperPoD and SuperCluster systems, which the company says can connect up to 256,000 accelerator cards in a single cluster. Huawei paired the announcement with a proprietary cluster interconnect architecture aimed at competing with Nvidia in China's domestic AI silicon market.
Key features
Commercial launch timeline moved up from late 2027 to Q1 2027
Companion Ascend 960PR variant announced alongside it
Anchors the Atlas 950 SuperPoD and SuperCluster systems
Atlas 950 SuperCluster scales to 256,000 accelerator cards
Paired with a proprietary domestic AI interconnect standard
Use cases
Large-scale domestic AI training/inference clusters in China
Nvidia-alternative accelerator deployments under export restrictions
Hyperscale interconnect architectures built on Huawei's own standard
Pricing: unknown — no commercial pricing published; chip is pre-launch as of September 2026.
A Kane CLI extension that turns product requirement documents into permanently-linked, evidence-backed test coverage.
Details and sources
TestMu AI launched the Assurance Lifecycle inside its Kane CLI terminal-native testing tool on September 15, 2026. Teams describe what a product must do, and Kane CLI designs tests that prove it, with each test permanently linked to the requirement it verifies and coverage measured from sealed evidence rather than self-reported pass/fail counts; the suite is reconciled automatically as the product changes. Kane CLI itself launched in April 2026 as a terminal-native browser automation tool, and the Assurance Lifecycle is a new command set built on top of it.
Key features
Converts requirement documents directly into test coverage
Permanently links each test to the specific requirement it verifies
Measures coverage from sealed evidence rather than self-reported results
Automatically reconciles the test suite as the product changes
Built as a command-set extension of the existing Kane CLI
Use cases
Answering 'what exactly is covered, and how do we know?' from a PRD
Requirement-to-test traceability for compliance-sensitive teams
Keeping test coverage current as products evolve without manual suite maintenance
Pricing: unknown — no standalone pricing published for the Assurance Lifecycle feature set.
Egnyte's business-context foundation that maps content, people, projects, and systems so AI agents get accurate context without redundant processing.
Details and sources
Egnyte launched the Context Layer on September 15, 2026, a foundation that maps relationships across a company's content, people, projects, business systems, and organizational knowledge and exposes that map to AI agents up front. Egnyte says this eliminates redundant per-query processing and reduces cost while improving accuracy, and the layer powers Egnyte's task-based agents across industry-specific workflows in financial services, life sciences, media and entertainment, and architecture/engineering/construction (AEC).
Key features
Maps relationships across content, people, projects, and business systems into a single context graph
Serves as shared context for Egnyte's task-based AI agents instead of per-query re-processing
Vertical workflow support for financial services, life sciences, media/entertainment, and AEC
Built on Egnyte's existing content platform and security/governance layer
Use cases
Grounding AI agents in a company's actual org/content structure before task execution
CorelDRAW's September 2026 update adding AI text-to-vector generation and enhanced PowerTRACE bitmap-to-vector tracing.
Details and sources
Corel announced a September 2026 update to CorelDRAW Graphics Suite adding AI-powered text-to-vector generation, described as 'Powered by Artist Intelligence,' alongside enhanced PowerTRACE bitmap-to-vector tracing and expanded CorelDRAW Web capabilities. Corel positions the update around faster AI-assisted ideation and file preparation for print shops, sign makers, apparel decorators, and other production designers.
Key features
AI text-to-vector generation directly inside CorelDRAW
Enhanced PowerTRACE bitmap-to-vector tracing
Expanded CorelDRAW Web capabilities
Performance improvements aimed at production design workflows
Use cases
Generating vector artwork from text prompts for print/sign production
Converting bitmap logos/artwork to clean vector output
Faster file prep for apparel and sign-making shops
Pricing: unknown — bundled into CorelDRAW Graphics Suite; no separate price disclosed for the AI features.
AI software-development platform whose 'Droid' agents plan, code, test, and document software; raised $200M at a $5B valuation in September 2026.
Details and sources
Factory raised $200 million in September 2026 at a $5 billion valuation, up from $1.5 billion five months earlier, in a round backed by Blackstone, Khosla Ventures, Sequoia Capital, NEA, and angel investors including Marc Benioff. Factory's platform lets developers describe tasks in natural language, which its 'Droid' agents break into sub-steps and execute inside isolated 'Droid Computers' (cloud or on-premises), with an intelligent router assigning each sub-step to the most cost-efficient available LLM. Additional components include a Readiness Report command that checks whether a repository is agent-compatible, AutoWiki for automatically generated and updated documentation, and built-in quality/security checks; Factory says it self-improves its agents using anonymized usage data from thousands of daily sessions, and counts teams at Nvidia, Adobe, and Palo Alto Networks among its users.
Key features
'Droid' agents break high-level task descriptions into sub-steps and execute them autonomously
Isolated 'Droid Computers' execution environments, deployable cloud or on-premises
Intelligent routing assigns each sub-step to the most cost-efficient available LLM
Readiness Report command audits repo compatibility with agent workflows
AutoWiki auto-generates and maintains code documentation
Self-improves agent behavior from anonymized usage data across sessions
Use cases
Enterprise teams delegating multi-step coding tasks to autonomous agents
Organizations needing on-premises agent execution for compliance reasons
Auto-maintained internal documentation via AutoWiki
Pricing: unknown — no public pricing disclosed as of this funding announcement.
xAI's speech-to-text model with diarization, timestamps, and smart turn detection for streaming or batch transcription.
Details and sources
Grok Voice Transcribe 2.0 is xAI's second-generation speech-to-text model, available via the Speech-to-Text API in batch ($0.10/hour) and streaming ($0.20/hour) modes. xAI says it ranks first for accuracy among 32 streaming models on the Artificial Analysis leaderboard and cuts multilingual word error rate on short phrases from 20.6% to 6.8% versus version 1.0, at the same price. It adds speaker diarization, word-level timestamps with confidence scores, up to 8 independent audio channels, key-term biasing, and automatic formatting of numbers, dates, and contact details.
Key features
Batch and streaming transcription modes
Speaker diarization and word-level timestamps with confidence scores at no extra cost
Up to 8 independent audio channels per request
Key-term biasing for up to 100 domain-specific terms
Smart turn detection for voice agents
Automatic language detection across multiple languages
Use cases
Voice agent turn-taking and conversation transcription
Multilingual call transcription with speaker separation
Domain-specific transcription (e.g., medical, legal) via key-term biasing
Real-time captioning and meeting transcription
Pricing: $0.10 per audio hour for batch transcription, $0.20 per audio hour for streaming; diarization and timestamps included at no extra cost.
TypeSafe AI's first System One model: a non-LLM decision engine returning typed, calibrated choices instead of generated text.
Details and sources
Jev is TypeSafe AI's first System One model, built to make fast structured decisions such as routing, classification, scoring, and extraction by returning typed probabilistic outputs instead of free-form text. TypeSafe says it matches existing LLMs on System One-style decision tasks while responding in 70-500ms, which it markets as up to 193x faster and 445x cheaper than comparable LLM calls, and every output ships with a calibrated confidence score. It launched in early access on September 22, 2026, at $0.042 per million input tokens with output tokens priced at zero.
Key features
Returns typed, structured decisions instead of generated text
70-500ms response time via parallel rather than sequential sampling
Calibrated confidence score attached to every output
Vendor claims it cannot produce hallucinated or type-invalid outputs
Vendor claims up to 193x faster and 445x cheaper than comparable LLMs on decision tasks
Use cases
Routing and classification inside application logic
Real-time scoring and extraction pipelines
Verification steps in agentic workflows
High-volume conditional logic that doesn't need generated prose
Pricing: Early access: $0.042 per million input tokens; output tokens free ('too cheap to meter'). No published enterprise tier yet.
MIT-licensed platform for self-hosted AI coworkers, each with its own sandboxed browser, files, and policy-checked tools.
Details and sources
OpenBot is CopilotKit's open-source, self-hostable platform for running AI agents as 'coworkers' that each get an isolated computer container with a dedicated browser, credentials, and workspace. It runs any AG-UI-compatible framework (LangGraph, Mastra, CrewAI, Pydantic AI, and others) behind a unified gateway that checks a CEL policy engine before every browser, file, or tool action and logs a full audit trail. It ships under the MIT license, reached v0.0.15 on September 22, 2026 with OAuth sign-in for model providers, and had 5,293 GitHub stars and 680 forks at time of writing.
Key features
Each bot runs in its own isolated computer container (browser, credentials, workspace)
CEL-based policy engine approves or blocks actions before execution, with full audit logging
Runs any AG-UI-compatible agent framework (LangGraph, Mastra, CrewAI, Pydantic AI)
Encrypted credential storage and RBAC with SAML/OIDC/Google/Microsoft/Okta identity integration
MCP server integration and custom React UI components for agent responses
OAuth sign-in for model providers (Google, xAI) as of v0.0.15, avoiding raw API keys
Use cases
Self-hosted internal AI assistants that need company-controlled data and credentials
Unbiased's blended model that runs several frontier and open-source models per request and returns the best synthesized answer.
Details and sources
Pareto is Unbiased's composite model, built by the Circuit & Chisel team, that runs multiple frontier and open-source LLMs in parallel on each request and dynamically synthesizes or selects the best result rather than relying on one set of weights. It launched publicly on September 17, 2026 after a roughly 33-hour free stealth preview on OpenRouter under the codename 'Union Alpha' starting September 16. The model accepts a 262,144-token context window, can output up to 131,072 tokens, supports text and image input with tool calling and JSON output, and is priced at $2.50 per million input tokens ($0.25 cached) and $7.50 per million output tokens.
Key features
Blended architecture: runs several frontier/open-source models per request and keeps the best answer
Locks to one model for the rest of a conversation to preserve prompt-caching and consistency
262,144-token input context window, up to 131,072-token output
Text and image input, tool calling, and JSON output support
Scored 74 on DeepSWE, 51 on Terminal-Bench 4.0, 78 on MMMU-Pro, 88 on ArXivMath per vendor benchmarks
Use cases
Coding and agentic workflows needing frontier-level reasoning without picking a single vendor's model
Research and math-heavy tasks (per ArXivMath benchmark)
General-purpose tasks where teams want automatic best-of-N model selection
Pricing: $2.50 per million input tokens ($0.25 cached), $7.50 per million output tokens; no published task-cost figure or ongoing free tier beyond the initial stealth-preview window.
Alibaba's Qwen team open-weight 7B image model unifying generation and editing with native transparent-PNG output, research-license only.
Details and sources
Qwen released Qwen-Image-2.1 on September 20, 2026, a 7-billion-parameter open-weight image model that unifies text-to-image generation and image editing in one system and writes transparent PNGs natively via a real alpha channel. It accepts up to 10 reference images in a single pass, and weights landed simultaneously on Hugging Face, ModelScope, and GitHub with day-zero support in ComfyUI, Diffusers, vLLM-Omni, and SGLang. The model ships under a research-only license, so it can be downloaded and tested but not shipped in a commercial product without a separate grant from Alibaba.
Key features
7B-parameter visual generation component
Unifies text-to-image generation and image editing in one model
Native alpha-channel support for transparent PNG output
Accepts up to 10 reference images per generation/edit pass
Day-zero support in ComfyUI, Diffusers, vLLM-Omni, and SGLang
Use cases
Generating or editing images with transparency for design/compositing workflows
Reference-guided image editing using multiple source images
Research on open-weight multimodal image generation (non-commercial)
Pricing: Free open-weight download for research use; research-only license prohibits commercial deployment without a separate license from Alibaba.
An MCP server that lets AI coding agents draft, publish, and measure product changelogs, with every post held for human approval.
Details and sources
ReleasePad launched an MCP server on September 14, 2026 that connects AI assistants such as Claude, Codex, and Grok to its changelog product via OAuth (no API keys). Through it, an agent can read what a team merged, draft a customer-facing release note in the team's voice, categorize and illustrate it, schedule it, and pull up analytics on how past posts performed; every draft defaults to unpublished until a person approves it. Published entries distribute to an in-app widget, a public changelog page, and an LLM-readable Markdown feed. The MCP server ships in ReleasePad's Pro plan at $35 per product per month with unlimited team members; a free tier can connect but with limited functionality.
Key features
OAuth connection with no API keys to manage
Agents draft, categorize, illustrate, and schedule changelog posts from chat
Draft-by-default workflow; nothing publishes without human approval
Analytics on prior post performance accessible to the agent
Distributes to in-app widget, public changelog page, and LLM-readable Markdown feed
One-click install for Cursor; connector setup for Claude.ai and Claude Code
Use cases
Drafting release notes directly from an AI coding agent's session after a merge
Keeping a public changelog and in-app widget in sync without manual writing
Giving agents read access to past changelog performance to inform new posts
Pricing: Included in ReleasePad Pro at $35/product/month, unlimited team members; free tier can connect but with restricted functionality.
Cloudflare open-sourced security-audit-skill under MIT in June 2026 and publicly detailed it in a blog post on September 17, 2026; by then it had reached 7,481 GitHub stars and 430 forks. It orchestrates isolated coding-agent runs through reconnaissance, coverage-led vulnerability hunting, candidate validation, structured output, and independent re-verification, on the premise that a model's first answer about a vulnerability is only a candidate finding, not a confirmed one. The repository ships one attack-family module per file, including memory corruption, prompt injection, HTTP request framing, and tenant isolation, and is the single-repo skill that seeded Cloudflare's larger internal vulnerability-discovery harness.
