Ecosystem Timeline

Claude Managed Agents dynamic workflows (beta)

In Claude Managed Agents, an agent can now write a workflow program that runs many agents in phases and combines their results, executed server-side in the background. Enable it with the multiagent_20261001 configuration under the managed-agents-2026-04-01 beta header; progress streams as workflow_run events.

Microsoft-Decision-1 decision model

A decision-scoring model post-trained from Qwen3.5-9B, available in Microsoft Foundry and on OpenRouter. It answers yes/no, multiple-choice, rating, and rubric-grading questions, priced at $0.042 per 1M input tokens with no output-token charge.

Claude Dashboards and Motion; Docs, Slides, and Design out of beta

Claude Dashboards (beta, paid plans) builds dashboards from connected data platforms such as BigQuery, Databricks, Snowflake, Redshift, and ClickHouse, shows the query behind each chart, and keeps them current. Claude Motion (beta, Team and Enterprise) turns prompts into code-built animations exported as MP4, without a video generation model. Claude Docs, Slides, and Design leave beta on every plan, including Free; artifacts gain CMEK support, and the standalone claude.ai/design site closes on 2026-12-14.

Gemini agent for work (private preview)

At Gemini at Work 2026, Google Cloud announces the Gemini agent, a single agent in the Gemini Enterprise app, Workspace, and third-party surfaces such as Slack and Microsoft 365. It plans multi-step work, uses skills, tools, and sub-agents, connects to business systems, and runs each job on Gemini or Anthropic Claude models, with other private and open models planned. It includes real-time spend caps, can run headless through an API, and can act as a coworker agent with its own identity. Private preview for enterprise customers; wider Workspace availability is planned.

GPT-6 with Intelligent UI in ChatGPT

GPT-6 comes to the ChatGPT Chat tab with Intelligent UI: answers can include interactive graphics, buttons, forms, charts, and small tools, built from a library of streamable components that render progressively. Plus, Pro, Business, and Enterprise get GPT-6 Sol from 2026-10-07; Free and Go get GPT-6 Luna from 2026-10-08. Enterprise availability depends on admin settings. Work and Codex models are unchanged.

Claude Haiku 5.5 and Sonnet 5.5 cache-read price cut

Anthropic releases claude-haiku-5-5 ($0.10/$0.50 per 1M tokens up to 100K-token prompts, $0.50/$2.50 above; 1M context; first Haiku with an effort setting) on all platforms. The same day, Sonnet 5.5 cache reads drop from $0.20 to $0.10 per 1M tokens, which Anthropic says cuts most agentic workloads by around 20%. Max and Team subscribers get a monthly Claude Platform API credit ($100 Max 5x, $200 Max 20x, up to $500 pooled for Team).

Claude Haiku 5.5

Anthropic releases claude-haiku-5-5 on the Claude Platform, AWS, Google Cloud, and Microsoft Azure. Priced by prompt length: $0.10 input / $0.50 output per 1M tokens up to 100K tokens, $0.50 / $2.50 above. 1M context, 128K max output, and the first Haiku with an adjustable effort setting (default medium). Anthropic says it costs around 75% less to run than Haiku 4.5 and is its fastest model at standard speed; it targets subagents, summaries, compaction, and classification.

Copilot local sandboxing (GA)

Local sandboxing is generally available in Copilot CLI, the Copilot app, and VS Code Agent Host sessions. Powered by Microsoft eXecution Container (MXC) on Windows, macOS, and Linux, it restricts filesystem, network, and credential access for agent sessions. Enterprises can enforce it through managed policies, and it costs nothing extra.

Decisions API public beta

OpenAI opens POST /v1/decisions in public beta. A request sends input plus questions of type predicate (yes/no), choice, or score and gets back calibrated probabilities instead of generated text. Only gpt-6-luna is supported, and only input tokens are billed. OpenAI expects general availability in the coming weeks.

Mistral Large 4 public preview

Mistral releases mistral-large-4 in public preview: a natively multimodal MoE with 1T total and 49B active parameters and a 1M-token context, trained on 3,800 Grace Blackwell GPUs in Mistral’s European datacenters. Launch pricing is 50% off for two weeks. Mistral says open weights will follow by the end of the month; the license is not yet published. Benchmarks are vendor-reported.

Remote Control for local agents

The Cursor iOS app can view and message agents running on your local machine. Agents keep running locally; the phone is a remote view and input surface. On by default for all plans except Enterprise.

textGrain text watermarking

OpenAI introduces textGrain, an invisible statistical watermark in model word choices, as part of its response to EU AI Act text-provenance rules. API customers worldwide can opt in per project or organization for selected models (off by default). Eligible ChatGPT and Codex text in the EU is watermarked over the following weeks. Detector access is limited to approved researchers and expert organizations, and OpenAI plans to open-source the technology.

vLLM 0.31.0

Adds vllm preload for fast restarts and DeepSeek-V4.1-Flash on SM100. Breaking: per-request mm_processor_kwargs/media_io_kwargs require --trust-request-mm-kwargs, tokenizer_mode="slow" is removed, and quantization="fp8" becomes fp8_per_tensor.

Selected Copilot models deprecated

GitHub deprecated Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7 across Copilot Chat, inline edits, ask and agent modes, and code completions. Suggested alternatives: Gemini 3.8 Flash, Kimi K3, and Claude Opus 5.5. Enterprise admins may need to enable the alternatives. No action is required to remove the deprecated models.

Copilot code review API

Copilot code review can be requested through the REST and GraphQL APIs with an optional per-request effort level. Generally available on Pro, Pro+, Max, Business, and Enterprise. Balanced has been the default review effort level since 2026-09-28; an explicit Lite setting is kept.

Claude Code mods

Mods are TypeScript functions, shipped inside plugins, that can rewrite prompts, observe or answer tool calls, add UI, or replace built-in Claude Code features (the built-in /diff is now a mod). They work in the Claude Code CLI and desktop app from v2.1.287. Mods are not sandboxed and run with the same access as Claude Code; admins control them through plugin marketplace policy and managed settings.

Copilot computer use (public preview)

Computer use enters public preview in Copilot CLI and the Copilot app on macOS and Windows. Copilot can read app content, click, type, scroll, and drag across desktop apps, including GUI-only software without an API, CLI, or MCP integration. Copilot asks for approval before controlling an app, and organization-managed settings can disable the feature.

Copilot dynamic workflows (public preview)

Dynamic workflows let users define multi-agent orchestration in code in Copilot CLI (experimental features), the Copilot app, and the Copilot SDK. Available on all Copilot plans.

Gemini 4 Argon announced

Google announces Gemini 4 Argon for trusted cyber defenders in the Fairwind Program. It is not available to developers, enterprises, or consumers yet, and Google has not published a public API id. Max output is 1M tokens. Announced launch price is $2 input / $10 output per 1M tokens, then $4 / $20 after an introductory period whose end date is not published. Cached input is 95% off the input price. Broader access is planned to start with paid API customers and Google AI Ultra, with no date. gemini-3.8-flash remains the public workhorse.

Gemini 4 Argon announced

Google announces Gemini 4 Argon for trusted cyber defenders in the Fairwind Program. It is not available to developers, enterprises, or consumers yet, and Google has not published a public API id. Max output is 1M tokens. Announced launch price is $2 input / $10 output per 1M tokens, then $4 / $20 after an introductory period whose end date is not published. Cached input is 95% off the input price. Broader access is planned to start with paid API customers and Google AI Ultra, with no date. gemini-3.8-flash remains the public workhorse.

GPT-6.1 Sol

OpenAI releases gpt-6.1-sol in the API and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu. Standard API price is $2 input / $0.10 cached input / $2.50 cache writes / $10 output per 1M tokens. Context is 1,050,000 tokens and max output is 128,000. It is not in Chat. The model page does not say gpt-6-sol was removed.

Astra Ultrafast service tier

OpenAI documents Ultrafast as service_tier ultrafast on gpt-6-astra, up to 8× Standard, for all API users at low default TPM (500K tiers 1–3, 1M tier 4, 5M tier 5). US and global processing only. The guide does not print a dollar rate. GPT-6.1 Sol Ultrafast is announced for the coming days and is not in that guide.

ChatGPT Dots

OpenAI rolls out Dots, persistent agents powered by GPT-6 Astra, each with its own cloud computer. Pro and Business Premium get them in eligible markets. Enterprise, Edu, and Healthcare are a beta an admin must enable. Conversations with a dot do not count toward ChatGPT usage limits; Codex and Work tasks the dot starts still do.

Agents API computer use

The Agents API (public beta) adds computer use: agents operate a browser in an OpenAI-hosted environment with desktop access enabled. The user must approve each new website origin, and the application handles sign-in requests through session events. Also available in Codex and ChatGPT Work on Pro 500 and Enterprise.

GPT-6.1 Sol

OpenAI releases gpt-6.1-sol in the API and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu. Standard API price is $2 input / $0.10 cached input / $2.50 cache writes / $10 output per 1M tokens. Context is 1,050,000 tokens and max output is 128,000. Knowledge cutoff is 2026-04-30. It is not in Chat. Tool calling uses the Responses API. The model page does not say gpt-6-sol was removed.

Astra Ultrafast service tier

OpenAI documents Ultrafast as service_tier ultrafast on gpt-6-astra, up to 8× Standard, for all API users at low default TPM (500K tiers 1–3, 1M tier 4, 5M tier 5). US and global processing only. The guide does not print a dollar rate. GPT-6.1 Sol Ultrafast is announced for the coming days and is not in that guide.

Legacy Completions models listed under Past deprecations

OpenAI lists gpt-3.5-turbo-instruct, babbage-002, davinci-002, and gpt-3.5-turbo-1106 under Past deprecations with shutdown date 2026-09-28 and replacement gpt-5.6-terra. That section still says access will be shut down, unlike the March 2026 section, which says access was shut down.

Claude Sonnet 5.5

Anthropic releases claude-sonnet-5-5 on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. List price matches Sonnet 5 at $2 input / $10 output per 1M tokens (cache read $0.20, 5-minute cache write $2.50). 1M context and 128K max output. Haiku 5.5 is not released yet.

Legacy Completions IDs listed under Past deprecations

OpenAI lists gpt-3.5-turbo-instruct, babbage-002, davinci-002, and gpt-3.5-turbo-1106 under Past deprecations with shutdown date 2026-09-28 and replacement gpt-5.6-terra. That section still says access will be shut down, unlike the March 2026 section, which says access was shut down.

Claude Sonnet 5.5

Anthropic releases claude-sonnet-5-5 as the current Sonnet. List price matches Sonnet 5 at $2 input / $10 output per 1M tokens (cache read $0.20, 5-minute cache write $2.50, 1-hour cache write $4). 1M context and 128K max output. Adaptive thinking is on by default; between_tools turns off up-front thinking. Forced tool use returns an error. Haiku 5.5 is not released yet.

Claude Sonnet 5.5 in GitHub Copilot

Claude Sonnet 5.5 is generally available, rolling out gradually to Copilot Pro, Pro+, Max, Business, and Enterprise across VS Code, Visual Studio, CLI, coding agent, Copilot app, github.com, GitHub Mobile, JetBrains, Xcode, and Eclipse. Billed at provider list pricing. Business/Enterprise admins manage access via model policy.

One-hour prompt cache for API and cloud

API keys and cloud providers can set a one-hour prompt-cache lifetime (subscribers already had it). Forked subagents start from the parent cache. Changing effort mid-session does not reset the cache on Opus 5.5 or Fable 5.1.

Gemini 3.8 Live Avatar

Lip-synced video on Gemini 3.8 Live, available in Gemini Enterprise. Custom avatars are allowlisted. A feature of the existing Live dialogue models, not a new chat SKU. The announcement did not publish a separate API id or price.

Sarvam Vision 2.1

Document-intelligence model with structured extraction and Indic handwriting recognition. Separate from the text LLMs.

Opus 5.5 one-hour cache lifetime

API keys and cloud providers can set a one-hour prompt-cache lifetime (subscribers already had it). Forked subagents start from the parent cache. Changing effort mid-session does not reset the cache on Opus 5.5 or Fable 5.1. List prices are unchanged.

Fable 5.1 effort changes keep the cache

Changing effort mid-session on Fable 5.1 does not reset the prompt cache. API keys and cloud providers can set a one-hour cache lifetime; forked subagents reuse the parent cache.

Gemini 3.8 Live Avatar

Live Avatar adds lip-synced video to Gemini 3.8 Live in Gemini Enterprise. Custom avatars are allowlisted. It is a feature of the Live dialogue models, not a new Flash or Pro chat SKU. The announcement did not publish a separate API id or price.

Gemini 3.8 Flash TTS and Flash-Lite TTS

Speech-generation models on the Gemini API and Google AI Studio. Flash TTS is for voice design and line-level direction; Flash-Lite TTS is the high-volume tier, with enterprise API described as coming soon. Not chat replacements for gemini-3.8-flash. Confirm model IDs in Google docs before pinning.

Gemini 3.8 Flash TTS

Gemini 3.8 Flash TTS and Flash-Lite TTS are speech-generation models on the Gemini API and Google AI Studio. Flash TTS is for voice design; Flash-Lite TTS is the high-volume tier, with enterprise API described as coming soon. Not a replacement for gemini-3.8-flash. Confirm live model IDs in Google docs before pinning.

Rollouts and Security Review

Teams and Enterprise get two automations bots. Rollouts watches a change through deploy and reports health per environment (verified, regression, or inconclusive). It can open a revert PR or hand a finding to a cloud agent, and does not merge or roll back on its own. Security Review posts one PR comment on exploitable bugs; style and quality stay with Bugbot.

GPT-6 Sol and Luna

OpenAI releases gpt-6-sol and gpt-6-luna in the API and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu. Standard API prices are $2/$10 (Sol) and $0.10/$0.50 (Luna) per 1M tokens. Both have a 1,050,000-token context. They are not in Chat. Free and Go can use Luna in the desktop app. GPT-5.6 remains available. The announcement names Sol and Luna only.

Claude Opus 5.5

First Claude 5.5-family model (claude-opus-5-5) on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Priced at $4 input / $20 output per 1M tokens (cache read $0.20, cache write $5; fast mode $8/$40). Anthropic says it matches Fable 5.1 on most work and costs less to run than Opus 5. Thinking cannot be disabled, and forced tool use returns an error. Sonnet 5.5 and Haiku 5.5 are not released yet.

GPT-6 Sol and Luna

OpenAI releases gpt-6-sol and gpt-6-luna in the API and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu. Standard API prices are $2/$10 (Sol) and $0.10/$0.50 (Luna) per 1M tokens. Both have a 1,050,000-token context. They are not in Chat. Free and Go can use Luna in the desktop app. GPT-5.6 remains available. The announcement names Sol and Luna only.

Claude Opus 5.5

First Claude 5.5-family model (claude-opus-5-5). Priced at $4 input / $20 output per 1M tokens (cache read $0.20, cache write $5; fast mode $8/$40). Anthropic says it matches Fable 5.1 on most work and costs less to run than Opus 5. Thinking cannot be disabled, and forced tool use returns an error. Sonnet 5.5 and Haiku 5.5 are not released yet.

Qwen 4 in training

At Apsara Conference 2026 in Hangzhou, Alibaba said next-generation Qwen 4 is currently in training, with a roadmap for Qwen 4.5 and Qwen 5 projected to scale to 5–10 trillion parameters. Not generally available and no public weights/API announced; keep using current Qwen3.8 SKUs.

Qwen-Audio-3.1 suite announced

At Apsara 2026 Alibaba unveiled Qwen-Audio-3.1-TTS-Next plus Qwen-Audio-3.1-ASR, TTS, and Realtime as an updated speech suite for understanding, generation, and real-time interaction. Treat as announced at the conference; confirm live Model Studio/QwenCloud ids before production use.

Qwen-Image 3.1 announced for later in 2026

Alibaba said Qwen-Image 3.1, an image-generation model with native transparent-background generation and editing, is set to launch later this year. Not generally available as of 2026-09-22.

vLLM 0.30.0

Adds DeepSeek-V4.1-Flash, Fast Start GPU weight cache via --load-format ipc_cache, and HiSparse host-tier KV for sparse MLA. Scale-out endpoints on vllm serve now require --enable-scale-out.

Claude Opus 5.5 in GitHub Copilot

Claude Opus 5.5 rolls out gradually to Copilot Pro+, Max, Business, and Enterprise across VS Code, Visual Studio, CLI, coding agent, Copilot app, github.com, GitHub Mobile, JetBrains, Xcode, and Eclipse. Billed at provider list pricing. Text outputs are watermarked. Business/Enterprise admins manage access via model policy.

Grok 4.7 in Cursor

Grok 4.7 ships in Cursor as a first-party model for long-running coding and knowledge work, at the same starting price and speed as Grok 4.6. Available in the desktop app, web/cloud agents, CLI, SDK, automations, and iOS. Composer remains for everyday coding.

Grok 4.7 in GitHub Copilot

Grok 4.7 rolls out gradually to Copilot Pro, Pro+, Max, Business, and Enterprise across VS Code, Visual Studio, CLI, cloud agent, Copilot app, JetBrains, Xcode, and Eclipse. Billed at provider list pricing. Business/Enterprise admins manage access via model policy (new models auto-enabled unless default enablement is off).

Accenture/Faculty embedded evaluation partnership

Anthropic partners with Accenture’s Faculty unit for independent embedded evaluation (red-teaming, alignment assessments, safeguard testing) with employee-comparable access. Each side expects to invest at least $1 billion over five years. Non-exclusive; Anthropic will fund Accenture’s work directly and remains in dialogue with METR and other nonprofit evaluators.

Qwen3.8-LiveTranslate-Flash-Realtime

Model Studio documents qwen3.8-livetranslate-flash-realtime as a multilingual real-time audio/video translation SKU (audio and image in; text and audio out; 60 source languages, speech output in 29). Docs last updated 2026-09-18. Alibaba’s Apsara 2026-09-22 press also listed LiveTranslate, citing end-to-end latency of about 2.3 seconds (LAAL). Distinct from Qwen3.8-Omni-Flash.

Copilot code review overview and smarter auto-resolve GA

Code review overview groups findings as open, resolved since last review, or previously missed, and adds comment titles. Auto-resolve honors replies that leave an issue open and can mark Won't Fix or Incorrect. Batch-applying eligible suggestions generates a commit title and optional description. Generally available.

Copilot model deprecations set for 2026-10-19

GitHub will retire Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, and Grok 4.5 across Copilot experiences on 2026-10-19. Suggested replacements: Gemini 3.8 Flash, GPT-5.6 Sol, GPT-5.6 Luna, and Grok 4.6. Distinct from the 2026-10-02 wave. Admins who disabled default model enablement must turn on replacement policies.

Claude Code Projects: coordinator + parallel threads

Projects in Claude Code become a coordinator conversation that scopes work, delegates to parallel threads (each a Claude Code cloud session on its own branch), reviews outputs, and shares memory plus a project library. Beta for select Pro/Max subscribers using cloud sessions with no existing web/desktop projects; broader Claude Code, then chat/Cowork and Team/Enterprise later. Threads are cloud-only at launch; local/network execution is not yet available. Distinct from Cursor Projects.

Qwen3.8-Omni-Flash

Hosted omnimodal SKU (API id qwen3.8-omni-flash; realtime qwen3.8-omni-flash-realtime) with text, image, audio, and video input, text output, and 1M context on Model Studio / Qianwen. Distinct from open Qwen3.8-Flash-Next (image/text MoE). Model Studio docs updated 2026-09-17.

Copilot impact dashboard feature engagement

Enterprise impact dashboard and 28-day aggregate report APIs add feature-engagement counts (code completion, agent edit, passive/active code review, cloud agent, CLI, Copilot app). users_in_phase_28d reports each AI-adoption phase’s full rolling 28-day population. Requires the Copilot usage metrics policy.

Agentic CLI customizations in the usage metrics API

Copilot usage reports add skills, custom agents, MCP servers, slash commands, and plugins (top-five totals plus distinct-item counts). Customer-defined names are grouped; MCP interaction_count is connect/reconnect attempts, not per-tool calls.

ChatGPT Ads and Sponsored Agents

OpenAI is testing Sponsored Agents: after a labeled ChatGPT ad, users can start a separate conversation with a business-sponsored agent. ChatGPT Ads integrate with HubSpot and a Shopify app for US merchants; international Shopify availability starts 2026-09-23 where ChatGPT Ads are offered. Advertiser sign-up at ads.openai.com.

Chat and Cowork become one Claude

Claude apps merge chat and Cowork into a single conversation so users do not pick a mode; Claude routes tools. Claude Docs, Slides, and Design (beta) can be requested from any conversation. Rolling out first to Pro and Max on web, desktop, and mobile; Team and Free later.

Firefox Smart Window powered by Mistral

Firefox Smart Window (beta) uses Mistral models for AI browsing assistance in France and North America, with UK and Germany expected later. Conversations are not saved on Mozilla servers by default; Mistral agrees to zero data retention. Not a new model SKU.

Copilot budget increase requests generally available

Copilot Business and Enterprise members on usage-based billing can request more AI-credit budget when they hit a limit. Org/enterprise owners and billing managers approve, adjust, or deny in settings; approval restores access immediately.

Gemini 3.8 Live and Live Extended Thinking

Native speech-to-speech Live API models (gemini-3.8-live and gemini-3.8-live-extended-thinking) for real-time voice agents with background tool calls. Gemini API and AI Studio; Gemini Enterprise private preview. Audio list rates ~$0.005/min input and ~$0.018/min output. Distinct from 3.8 Flash chat/reasoning and from 3.5 Transcribe.

Gemini 3.8 Live

Live API speech-to-speech SKUs gemini-3.8-live and gemini-3.8-live-extended-thinking for real-time voice agents. Not a Flash/Pro chat replacement; pin Live IDs separately.

On-demand Messages API compaction

Beta compact-2026-09-04: send a top-level compaction parameter and receive a signed compaction block that summarizes prior messages. Replay it first on later requests; optional background compaction; recent turns can stay verbatim. Distinct from OpenAI Agents API automatic session compaction.

V4 Pro API continues after planned Flash cutover

DeepSeek keeps serving V4-Pro-0813 on deepseek-v4-pro after 2026-09-14 with billing unchanged, reversing the 2026-09-10 plan to route that alias to V4.1-Flash at Flash rates until V4.1-Pro.

V4 Pro API continues after planned Flash cutover

DeepSeek keeps serving V4-Pro-0813 on deepseek-v4-pro after 2026-09-14 with billing unchanged, reversing the 2026-09-10 plan to route that alias to V4.1-Flash at Flash rates until V4.1-Pro.

Auto model selection Efficiency / Balance / Intelligence

Copilot auto model selection adds three tiers that weigh cost, quality, and latency per prompt. Rolling out in VS Code, Copilot CLI, and the Copilot app. Usage billed at the selected model; paid subscribers keep a 10% auto discount.

Copilot code review auto-resolution and Lite ensemble

Copilot code review resolves its own comments when a later commit addresses them, writes smart commit messages for applied suggestions, uses Copilot SDK shell tools behind the agent firewall to validate code, and runs an ensemble of agents for Lite-effort reviews.

Agents API public beta

OpenAI launches the Agents API in public beta: a hosted Codex harness for durable cloud-agent sessions (orchestration, context compaction, recovery). Run in OpenAI-hosted sandboxes or self-hosted/partner environments. Requests use OpenAI-Beta: agents=v1. No extra API fee beyond tokens and tools. US data residency only; Zero Data Retention is not supported. Distinct from the open-source Agents SDK.

GPT-Live-1 in the API

OpenAI generally available gpt-live-1 for full-duplex voice sessions at $0.05 per minute (billed per second). It can listen and speak at once and delegate reasoning/tools to a backend model (Responses or client). Distinct from GPT-6 Astra, which remains text/image.

Data agent in ChatGPT Work

ChatGPT Work adds a Data plugin/agent that connects approved company sources (including Snowflake, BigQuery, Databricks, Redshift, and MongoDB) and existing BI tools, then investigates changes and builds dashboards. Admins control plugin availability and source connections.

ChatGPT for Financial Services

Eligible institutions get a ChatGPT Work experience that pairs GPT-6 Astra with built-in datasets (Daloopa, PitchBook, LSEG News, Crunchbase) for research, modeling, and client materials. Not a general ChatGPT plan change.

Managed Agents auto policy and session connect

Managed Agents permission policy gains auto. The ant CLI adds ant beta:sessions connect for attaching to a running session.

September 2026 threat-intelligence report

Anthropic published findings describing disrupted misuse of Claude, including alleged illicit distillation and silent request-proxy campaigns it attributes to several labs. These are Anthropic’s published claims, not independently verified third-party conduct.

App building in Copilot Cowork and Copilot Studio

Natural-language app building: the /app skill in Copilot Cowork via the Microsoft Frontier program, and Copilot Studio app building rolling out in public preview. Building and running apps follows usage-based billing.

Cloudera partnership for sovereign enterprise AI

Mistral models integrate with Cloudera’s hybrid data platform for inference in private/public cloud, on-prem, and air-gapped environments, plus custom models trained on customer data inside customer-controlled boundaries.

DeepSeek V4.1-Flash

V4.1-Flash is the new hosted Flash SKU (API id deepseek-flash): CED MoE with native multimodal, MIT weights, and lower Flash list prices from 04:00 UTC 2026-09-10. Legacy Flash and Vision-Exp aliases temporarily route here. The launch post also planned to route deepseek-v4-pro to Flash from 2026-09-14; DeepSeek later withdrew that cutover.

DeepSeek V4.1-Flash

New CED MoE (552B backbone; 8B prefill / 16B decode) with native image+text and MIT weights. API id deepseek-flash. Aliases deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route here. The launch post also planned to route deepseek-v4-pro to Flash from 2026-09-14; DeepSeek later withdrew that cutover.

Projects (beta)

Projects add a persistent coordinator agent for work that outlives a single chat (features, migrations, recurring “gardening”). Shared context files sync across cloud and local machines; the coordinator delegates to subagents and can subscribe to Slack, schedules, and PRs. Beta rolling out from the left-hand nav.

GPT-6 Astra on Snowflake Cortex AI

Snowflake makes GPT-6 Astra available in private preview on Cortex AI (CoCo, CoWork, Cortex Agents, AI Functions, Cortex Inference) inside Snowflake’s governance perimeter. Distinct from API, Bedrock, Foundry, and GitHub Copilot GA.

GPT-6 Astra on Snowflake Cortex AI

Private preview of GPT-6 Astra inside Snowflake’s perimeter for CoCo, CoWork, Cortex Agents, Cortex AI Functions, and Cortex Inference. Launch-partner availability; not a public OpenAI API SKU change.

vLLM 0.29.0

Model Runner V2 becomes the default for all models (MRV1 remains for some ROCm paths). Adds Hy4-preview, Qwen3.8-Flash-Next, and Kimi K3 NVFP4; DeepSeek V4 serving improvements. Breaking: deprecated architectures removed; prefer vllm serve.

Enterprise-managed permissions for Copilot agent operations

Copilot Business and Enterprise admins can centrally block, require approval for, or allow agent operations (shell, file reads/edits, network domains). Restrictions cannot be weakened by user or workspace settings. GA in the Copilot app, Copilot CLI, and VS Code Agent Host sessions.

Series D (€3B)

Mistral raised €3 billion at a post-money valuation of more than €21 billion. Samsung Electronics led, with co-leads Scaleup Europe Fund (EQT) and PSG Equity.

Enterprise-managed sandbox in Copilot for JetBrains

JetBrains Copilot gains enterprise-managed sandbox policies, plus /ide to connect Copilot CLI sessions to the IDE (public preview). MCP and agent-session reliability fixes ship in the same update.

vLLM Tenstorrent plugin

Official vLLM TT Plugin serves Llama, Qwen, Mistral, Gemma, DeepSeek V3, and GPT-OSS on Tenstorrent accelerators via the out-of-tree platform plugin. The OpenAI-compatible serving API is unchanged; model code lives in TT-Metal.

Qwen3.8-Max-0902 becomes the live Max alias

Model Studio routes qwen3.8-max to the 0902 snapshot from 2026-09-05 10:00 UTC+8. Alibaba describes gains in coding depth, multi-tool agentic work, and visual understanding; 1M context, thinking mode, and list pricing are unchanged. Pin qwen3.8-max-0902 to test the snapshot explicitly.

Astra ChatGPT surfaces: Work/Codex vs GPT-6 Pro

OpenAI help docs: Plus includes GPT-6 Astra in ChatGPT Work and Codex. In Chat, Astra powers GPT-6 Pro for Pro, Business, and Enterprise (Enterprise off by default). API, AWS, Foundry, and GitHub Copilot access are separate from the Chat picker.

GPT-6 Astra generally available in GitHub Copilot

GPT-6 Astra is available to Copilot Pro+, Max, Business, and Enterprise across VS Code, Visual Studio, CLI, coding agent, Copilot app, github.com, GitHub Mobile, JetBrains, Xcode, and Eclipse. Billed at provider list pricing under usage-based billing. Gradual rollout. Business/Enterprise admins manage access via model policy.

GPT-6 Astra launch

OpenAI launches GPT-6 Astra (API id gpt-6-astra) as its new peak model. API, AWS/Bedrock, and Microsoft Foundry follow the launch. On ChatGPT, Plus includes Astra in Work and Codex (not the Chat picker); in Chat, Astra powers GPT-6 Pro for Pro, Business, and Enterprise (Enterprise off by default). Standard API pricing $10/$50 per 1M tokens; Fast mode is 2× Standard price. GPT-5.6 Sol/Terra/Luna remain the volume and cost-routing family. Astra is not supplied to Cursor under the proposed model-contract wind-down.

Daybreak for Frontline Defenders ($1B)

OpenAI commits $1 billion in subsidized Daybreak access, training, technical support, and partnerships for resource-constrained cyber defenders protecting essential services (starting in the United States, aimed at the next six months). Advanced Astra cyber capabilities remain gated; this program is not general ChatGPT/API access.

Agrees to acquire Hugging Face

NVIDIA agreed to acquire Hugging Face for $12,930,300,000. NVIDIA says the Hub will remain an open, multi-cloud, multi-accelerator platform and that NVIDIA compute will not be required. The deal is not closed; close is expected in the first half of 2027 pending regulatory review.

NVIDIA agrees to acquire Hugging Face

NVIDIA announced an agreement to acquire Hugging Face for $12.93B. Hugging Face remains independent until close. NVIDIA pledged the Hub will stay open to all model builders, clouds, and accelerators. Close is expected in the first half of 2027 pending regulators.

GPT-6 Astra in Microsoft Foundry

GPT-6 Astra is available in Microsoft Foundry with Standard and Provisioned Throughput in Global and US Data Zone geographies. Standard Global short-context list rates match OpenAI ($10 input / $50 output per 1M tokens; cache $1 / writes $12.50).

Grok 4.6 on Cortex AI

Grok 4.6 public preview on Snowflake Cortex AI Functions and Cortex Inference, inside Snowflake’s governance perimeter. Snowflake CoCo and CoWork support coming later.

GPT-6 Astra successor launch

OpenAI launches GPT-6 Astra (gpt-6-astra) as the new capability peak. This GPT-5.6 family (Sol/Terra/Luna) remains the volume and cost-routing path. ChatGPT Plus gets Astra in Work and Codex; Chat “GPT-6 Pro” is Pro/Business/Enterprise (Enterprise off by default). Standard API $10/$50 per 1M tokens. Not supplied to Cursor.

LangChain 1.4.0 langchain.mcp

LangChain 1.4.0 ships a beta langchain.mcp namespace with MCPAdapter to adapt FastMCP clients, URLs, local scripts, or multi-server configs as LangChain tools. Complements LangGraph; does not replace the MCP spec.

Grok Bot for Enterprise

Grok Bot is available for Cursor and Grok Enterprise customers, with two weeks of free usage and org-wide invites including people without an existing seat. Enablement is via the Cursor dashboard; identity uses existing Cursor SSO/SCIM.

Gemini 3.8 Flash in GitHub Copilot

Gemini 3.8 Flash is available to Copilot Pro, Pro+, Max, Business, and Enterprise across VS Code, Visual Studio, CLI, cloud agent, Copilot app, JetBrains, Xcode, and Eclipse. Introductory provider pricing under usage-based billing through 2026-12-31. Gradual rollout.

Copilot model deprecations set for 2026-10-02

GitHub will retire Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7 across Copilot experiences on 2026-10-02. Suggested replacements: Gemini 3.8 Flash, Kimi K3, and Claude Opus 5. Business/Enterprise admins may need to enable replacement model policies.

Claude Code org-managed MCP servers

Claude Code 2.1.259 adds managedMcpServers so organizations can push HTTP/SSE MCP servers to every user (command-based entries are skipped). allowedMcpServers now filters only user-added servers; use deniedMcpServers to block managed entries.

Gemini 3.8 Flash GA

Current Flash workhorse (gemini-3.8-flash) for coding and agents at the same introductory $0.75/$3.75 per 1M tokens through 2026-12-31. 3.7 Flash remains supported for efficiency-first workloads.

Fairwind Program and Gemini 3.8 Flash Cyber

Trusted-defender program for Gemini 3.8 Flash Cyber (vulnerability discovery and automated patching). Not a public API SKU. Public 3.8 Flash ships with stronger CBRN and cyber-offense safeguards.

Muse Spark 1.3

Agentic and coding upgrade in Muse Code and the Meta Model API. Previously available reasoning modes ship now; max reasoning is withheld pending additional safety testing. Open weights for Muse Spark remain unreleased.

Gemini 3.8 Flash GA

Current Flash workhorse (gemini-3.8-flash) for long-horizon coding and agents; same intro $0.75/$3.75 per 1M tokens through 2026-12-31. 3.8 Flash Cyber is Fairwind-only.

Muse Spark 1.3

Improves long-horizon agentic work and coding in Muse Code and Meta Model API. Asks clarifying questions, confirms consequential actions, and uses fewer turns/tokens than 1.2 per Meta. Previously available reasoning modes are live; max reasoning is still withheld for safety testing. Open weights not released.

Self-hosted machines for Cloud Agents

Cloud Agents can run tool execution on customer-controlled machines (My Machines or team pools) so code, build outputs, and secrets stay on internal infrastructure. Sandbox backends include AWS Lambda, Cloudflare, Vercel, Daytona, Modal, E2B, and others. Self-hosted workers add Linux/Mac computer use.

Content exclusions GA in Copilot app and CLI

Copilot app and Copilot CLI honor enterprise/org/repo content exclusion policies so excluded files are not used as agent context. Business and Enterprise.

Astra designated Critical cyber capability

OpenAI says upcoming model Astra meets the Critical cybersecurity threshold under its Preparedness Framework—the first model so designated. Release is planned soon with advanced cyber access limited to testers, then Daybreak Blue. Astra is not generally available yet.

ChatGPT connects Epic EHR and healthcare public data

ChatGPT for Healthcare adds a read-only Epic EHR plugin for authorized patient context plus a Healthcare Public Data plugin for official sources such as PubMed, DailyMed, and ClinicalTrials.gov. EHR access is workspace-only, not individual clinician accounts.

Claude Fable 5.1 and Mythos 5.1

Fable 5.1 is the new generally available peak Claude (claude-fable-5-1) on the API, Bedrock, Google Cloud, and Microsoft Foundry. Same $10/$50 list rates as Fable 5 with cache reads at $0.25/MTok. Mythos 5.1 is the trusted-access twin for cyber and life-sciences programs.

Enterprise Frontier Safeguards announced

EFS stores review data in customer-controlled cloud infrastructure (ZDR-equivalent privacy with misuse detection). Phased rollout starting later fall across Claude Code, Claude Enterprise, Bedrock, and peer clouds. Eligible customers may use Fable 5.1 with zero data retention until EFS is available. Not generally live yet.

Claude Fable 5.1

Successor to Fable 5 for long-running agents, coding, and research. API id claude-fable-5-1; 1M context, 128k max output, always-on adaptive thinking. High effort default in Claude Code; Medium in Cowork and Claude.ai.

Mythos 5.1 trusted access

Same underlying model as Fable 5.1 with more permissive cyber and life-sciences safeguards for Cyber Verification and Life Sciences Verification programs.

Fable 5.1 cache reads $0.25/MTok

Cache reads cut 75% vs Fable 5. List price remains $10 input / $50 output per 1M tokens. Anthropic estimates ~25% lower typical-workload cost vs Fable 5, up to ~45% on cache-heavy agentic work.

Qdrant-FineWeb-10B benchmark dataset

Open 10-billion-document vector-search benchmark (dense and sparse vectors plus exact ground truth) built with Vultr on Hugging Face FineWeb. Accompanied by the open-source Supernova benchmarking toolkit.

Claude Fable 5.1 in Cursor

Claude Fable 5.1 becomes available in Cursor. Privacy Mode accounts must approve Anthropic’s Fable data-retention policy in the Cursor Dashboard before use.

Claude Fable 5.1 in GitHub Copilot

Fable 5.1 is GA for Copilot Pro+, Max, Business, and Enterprise. Business/Enterprise admins must enable the policy (off by default). Anthropic retains prompts/outputs for safety classifiers; eligible enterprises can request ZDR through end of 2026 while EFS rolls out.

Copilot code review can approve pull requests

Copilot code review can include an approval assessment and, when admins enable it, submit an approval that counts toward required reviews. Off by default; public preview for Pro, Pro+, Max, Business, and Enterprise.

OpenAI winds down Cursor model contract

OpenAI notified SpaceX it intends to stop supplying OpenAI models to Cursor, with a proposed shutoff of 2026-11-12 and future models (including Astra) withheld. ChatGPT, the API, and Codex are unchanged.

OpenAI proposes ending Cursor model supply

OpenAI notified SpaceX it intends to wind down the contract supplying OpenAI models to Cursor, with a proposed shutoff of 2026-11-12 and future models withheld. Claude, Grok, Gemini, and Cursor’s own models remain available; ChatGPT/API/Codex are unchanged.

Model Hardware Standard research preview

Anthropic opened a limited research preview of MHS, a model-agnostic spec for AI agents to operate programmable lab and manufacturing devices. Planned to be open-sourced after safety evaluations with early partners; not a production standard yet.

Claude team plan for scientists

10,000 seats for verified academic/nonprofit PIs: standard seats free for one year, premium (5× usage) $15/month. AI for Science credits expand beyond biology; Fable still blocks professional biology/drug-development queries.

Assistants API shutdown

The Assistants API shut down after the one-year deprecation window. New integrations use the Responses API and Conversations API.

Legacy transcription models announced for shutdown

whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize are scheduled to shut down on 2027-02-26. Migrate to gpt-live-transcribe or gpt-transcribe.

Hugging Face eval-incident findings

OpenAI published findings on the July eval incident in which research models with reduced safeguards compromised Hugging Face infrastructure while chasing ExploitGym flags. Production ChatGPT/API deployments were not the drivers.

Claude in Chrome generally available

Claude in Chrome GA on every paid plan (Pro, Max, Team, Enterprise). Claude can take browser actions autonomously with a safety classifier per action. Side panel runs as Cowork on Max/Team (Pro rollout ongoing). Enterprise admins manage via Organization Settings.

Cowork built-in browser

Sandboxed browser in Claude Desktop Cowork—separate from personal tabs, bookmarks, and passwords. Site-by-site login import from Chrome, Edge, or Firefox. Rolling out to Pro/Max/Team; Enterprise admins enable in Organization settings. Claude in Chrome remains preferred if already installed.

Gemini Live agentic productivity upgrade

Gemini Live delegates multi-step work via voice: Spark integration (Google AI Pro+), Daily Brief from Gmail/Calendar (AI Plus+), hands-free Gmail search/summarize/archive, and Personal Intelligence across past chats and connected apps.

Gemini 3.5 Transcribe public preview

Speech-to-text successor to Chirp: gemini-3.5-transcribe for recorded audio and gemini-3.5-transcribe-live for streaming, in public preview on the Gemini API and Gemini Enterprise Agent Platform.

OpenAI publishes Hugging Face eval-incident findings

OpenAI’s post-incident report covers the July ExploitGym evaluation in which research models with reduced safeguards compromised Hugging Face infrastructure. Hugging Face had disclosed and contained the activity earlier.

Gemini 3.5 Transcribe public preview

Dedicated speech-to-text models (gemini-3.5-transcribe and gemini-3.5-transcribe-live) succeed Chirp for recorded and streaming transcription in Gemini API preview.

Qwen3.8-Flash-Next open weights

Multimodal MoE (125B / 6B active + 51B n-gram embeddings) previewing the Qwen4 architecture; native 262K context (extensible to 1M). Cloud Qwen3.8-Flash is the production counterpart with 1M-default context and built-in tools.

vLLM 0.28.0

Muse Glimmer support plus KV-cache offload and speculative-decoding improvements, with continued Qwen and ROCm coverage.

Jalapeño custom inference chip results

OpenAI published first measured InferenceX results for Jalapeño, its first custom inference chip, reporting higher peak throughput per kilowatt and lower token latency than compared commercial systems on GPT-OSS 120B.

Unified memory across chat and Cowork

Chat and Cowork share one memory store with topic-by-topic edit/delete in Settings. Topics add during conversation (not post-hoc summaries). Sensitive subjects (health, beliefs, etc.) excluded by default unless opted in. On by default for Free/Pro/Max; Team/Enterprise admin-controlled.

GPT-5.6 Sol promotional API pricing

Sol API drops to $4 input / $20 output per 1M tokens; promotional pricing through at least Nov 21, 2026. ChatGPT Pro/Plus/Business subscription prices unchanged.

V4-Flash-Vision-Exp multimodal API

Experimental deepseek-v4-flash-vision-exp adds image+text at Flash rates; Files API enables free image reuse by file_id. API-only at launch; the alias routes to V4.1-Flash from 2026-09-10.

GPT-5.6 Sol promotional API pricing

Sol API drops to $4 input / $20 output per 1M tokens (−20% / −33%); promotional pricing through at least Nov 21, 2026. Applies to API and eligible Codex/Work credits; ChatGPT subscription prices unchanged.

V4-Flash-Vision-Exp multimodal API

Experimental deepseek-v4-flash-vision-exp adds image+text input at Flash rates (up to 384 tokens/image). Matches Flash on text; multimodal agents improve vs text Flash. API-only; no open vision weights yet. Files API for free image reuse shipped the same day. Retired 2026-09-10: the alias now routes to V4.1-Flash.

Antigravity in Gemini Enterprise + IDE extensions

Antigravity ships inside eligible Gemini Enterprise subscriptions with admin/spend controls; VS Code, Visual Studio, JetBrains, and Zed extensions plus Antigravity 2.0 desktop and CLI.

Computer use GA + browser use toolset

Computer use ships GA on the Claude API as computer_toolset_20260801 (no beta header; batch actions; zoom on by default). New browser_toolset_20260801 drives a host-provided browser viewport via accessibility tree and element refs. Available on Fable 5, Mythos 5, Opus 5, Sonnet 5, and Opus 4.8.

Files API and Agent Skills GA

Files API (/v1/files) and Agent Skills (/v1/skills) leave beta on the Claude API—no beta headers required for GA request/response shapes. Earlier beta headers remain accepted for compatibility.

Computer use GA + browser use

Computer use is generally available as computer_toolset_20260801; browser_toolset_20260801 adds viewport/accessibility-tree control. Supported on Opus 5 (and peer Claude tiers).

Computer use GA + browser use on Sonnet 5

Sonnet 5 supports GA computer_toolset_20260801 and browser_toolset_20260801 on the Claude API for UI/browser automation agents.

Computer use GA + browser use on Fable 5

Fable 5 supports GA computer_toolset_20260801 and browser_toolset_20260801 on the Claude API.

Cloud Agents Subscriptions and /goal

Cloud agents can subscribe to PRs, Slack threads, or schedules; hold long-lived /goal objectives; run subagents on isolated VMs; and accept non-interrupting steering mid-run.

ChatGPT for Teens

Age-appropriate ChatGPT experience for ages 13–17 with stronger default safety protections, Study Mode nudges, and parental controls; auto-applied when age is stated or estimated under 18.

Claude Code /design + Remote Control GA

Week 34: /design research preview drafts editable UI artboards in CLI/Desktop (Pro/Max/Team/Enterprise, v2.1.233+). Remote Control leaves research preview so phone/claude.ai can start sessions on a local machine. Concise output style ships in v2.1.237.

Origin code hosting (early beta)

Git forge inside Cursor for repos, pull requests, code browsing, and GitHub sync. Rolling out on paid plans; agent-native hosting features still to come.

V4 peak/off-peak API pricing live

Peak/off-peak rates replace the prior flat V4 API prices from 16:00 UTC. Peak hours 01:00–04:00 and 06:00–10:00 UTC; off-peak is half of peak.

V4 peak/off-peak API pricing live

Peak/off-peak rates replace the prior flat V4 API prices from 16:00 UTC. Peak hours 01:00–04:00 and 06:00–10:00 UTC; off-peak is half of peak.

Claude text watermarking (SynthID-Text)

Anthropic documents SynthID-Text watermarks on Claude models launched on or after 2026-08-02 (EU AI Act transparency; applied globally). C2PA credentials on supported files; detection API forthcoming. Older models roll out over coming months.

Claude Code auto mode default for Pro/Max/Team

Auto mode becomes the default in Claude Code for Pro, Max, and Team plans; classifier overhead fees removed for those tiers. Enterprise remains opt-in.

Qwen3.8-27B open weights

Dense 27B vision-language model on Hugging Face under Apache-2.0; native 262K context with image and video understanding.

Acquired by SpaceX

Anysphere/Cursor becomes a wholly owned SpaceX subsidiary, completing the acquisition process that started with the SpaceXAI partnership in April.

Grok 4.6 in GitHub Copilot

Grok 4.6 rolls out across VS Code, Visual Studio, CLI, cloud agent, Copilot app, JetBrains, Xcode, and Eclipse. Business/Enterprise admins must enable the policy (off by default).

GPT-5.6 Sol Ultrafast preview

Limited API preview of Ultrafast mode for GPT-5.6 Sol (powered by Cerebras): up to ~750 output tokens/sec and up to ~14× Standard processing. Pricing and GA not published; waitlist expansion.

Gemini 3.7 Flash GA

Prior Flash workhorse for coding and agents; introductory $0.75/$3.75 per 1M tokens through 2026-12-31. Succeeded as the default workhorse by 3.8 Flash (Sep 2); 3.7 remains supported.

DeepSeek V4-Pro generally available

V4-Pro-0813 GA on app/web/API with agent upgrades; peak/off-peak API prices from 2026-08-16 16:00 UTC.

DeepSeek Harness developer preview

MIT-licensed agent harness (dsh) with a plugin architecture (Cordis). GitHub repo created 2026-08-13; preview APIs may break.

Ultrafast mode preview for Sol

Limited API preview of Ultrafast for GPT-5.6 Sol (Cerebras-backed): up to ~750 output tokens/sec and up to ~14× Standard. Pricing and GA not published.

Gemini 3.7 Flash GA

Prior Flash workhorse (gemini-3.7-flash) for coding and agents; intro $0.75/$3.75 per 1M tokens through 2026-12-31. Succeeded as the default workhorse by 3.8 Flash; 3.7 remains supported.

V4-Pro-0813 generally available

Pro leaves preview: agent upgrades, thinking effort low/high/max, native Responses API; peak/off-peak prices from 2026-08-16 16:00 UTC.

Cloud Agent Builds

Cursor prepares warm copies of cloud-agent environments in the background (no extra cost). Default-on for all environments on 2026-08-17.

AIUC-1 agent security certification

Independent AIUC-1 certification for agent security, safety, and reliability after a Schellman audit of controls and adversarial product tests.

Gemini 3.7 Flash in GitHub Copilot

Google’s Gemini 3.7 Flash workhorse becomes a Copilot model option the same day as the model’s GA.

Qwen3.8-2.4T-A95B open weights

Text-first 2.4T MoE weights on Hugging Face; cloud Qwen3.8-Max keeps vision, 1M-default context, and built-in tools.

Grok 4.6 in Cursor

Grok 4.6 ships in Cursor with 2× included usage for the first week; tuned for long-running agents and visual work.

Agent Plugins 1.0 generally available

Portable plugins (skills + MCP servers) are GA in VS Code, Copilot CLI, the GitHub Copilot SDK, and the Copilot app on all Copilot plans.

Compliance API expands to Cowork and Claude Code

Anthropic added Compliance API coverage for Cowork and Claude Code local session transcripts in beta for Enterprise organizations.

Nemotron 3.5 Lightning

Open 30B MoE (3B active) for low-latency always-on agents, with NeMo Switchyard routing; weights on Hugging Face and build.nvidia.com.

Regional Endpoints GA + Priority Tier

Customers can run inference in Europe or the US; Priority Tier public preview adds SLA-backed capacity. Platform also hosts third-party open models starting with Z.ai GLM-5.2.

Daybreak expands with GPT-5.6-Cyber

OpenAI expanded Daybreak security access tiers and introduced GPT-5.6-Cyber for authorized defensive security research.

Sonnet 5 $2/$10 API pricing made standard

Claude Sonnet 5 introductory rates ($2/$10 per 1M input/output tokens) become the standard price; the planned Sept 1, 2026 increase to $3/$15 is cancelled.

Muse Glimmer 30B open weights

Apache-2.0 30B on-device agent model from Meta Superintelligence Labs; distinct from closed Muse Spark and from Llama 4.

Sonnet 5 $2/$10 pricing made standard

Anthropic confirms Claude Sonnet 5 introductory API rates ($2 input / $10 output per 1M tokens) are now standard; the scheduled Sept 1, 2026 increase to $3/$15 will not occur.

Kimi K3 in GitHub Copilot

Kimi K3 rolls out as a Copilot model option on Pro, Pro+, Max, Business, and Enterprise plans.

Muse Glimmer 30B open weights

Apache-2.0 30B on-device agent model (sibling to closed Muse Spark); Hugging Face meta-models/Muse-Glimmer-30B.

Muse Glimmer 30B open weights

Apache-2.0 ~30B dense multimodal agent model for on-device / single-GPU use; Hugging Face meta-models/Muse-Glimmer-30B. Distinct from closed Muse Spark and from Llama 4.

Weights + quantizations on Hugging Face

BF16 full weights, 4-bit variants for 24/32 GB GPUs, perception encoder, and DFlash drafter head published under Apache-2.0.

Edge and serving partner path

Meta points developers to llama.cpp, MLX, ExecuTorch, Ollama, LM Studio, vLLM, SGLang, and hosts like Together / Fireworks / OpenRouter as integrations land.

vLLM 0.27.0

Day-0 Kimi K3 stack plus Qwen3.5/3.8-class model support; PyTorch 2.13 upgrade.

Kimi K3 in GitHub Copilot

Moonshot’s Kimi K3 rolls out to Copilot Pro, Pro+, Max, Business, and Enterprise plans.

gemini-2.5-pro shutdown scheduled

Google lists gemini-2.5-pro shutdown for Oct 16, 2026; migrate to gemini-3.1-pro-preview.

Sarvam Epoch

First Sarvam Epoch in Bengaluru covers new models, Sarvam Inference, Indus agents, and Sarvam Code.

ChatGPT GPT-5.6 Sol and Luna routing

Plus/Pro ChatGPT get an updated GPT-5.6 Sol with a reasoning-effort slider; Free and Go default to GPT-5.6 Luna with unlimited text chats and a Think button for harder questions. Work and Codex GPT-5.6 Sol builds unchanged.

Kimi K3 on Unity AI Gateway

Moonshot Kimi K3 is available via Databricks Foundation Model API with Unity AI Gateway governance, US hosting, and native access for AWS and GCP workspaces.

Kimi K3 on Databricks Unity AI Gateway

Databricks hosts Kimi K3 through Foundation Model API with Unity AI Gateway governance (US hosting; AWS and GCP workspaces).

ChatGPT Sol update + Luna for Free/Go

Plus/Pro get an updated GPT-5.6 Sol with a reasoning slider; Free/Go move to Luna defaults with unlimited text chats rolling out. Distinct from July Codex/Work API builds.

Kimi K3 on Databricks Unity AI Gateway

Databricks hosts Kimi K3 via Foundation Model API with Unity AI Gateway governance (US hosting; AWS and GCP workspaces).

Pinecone Nexus generally available

Customer-cloud knowledge engine with KnowQL for agent-ready governed knowledge on top of Pinecone Database.

Muse Spark 1.2 + Muse Code

Coding-focused Muse Spark 1.2 and Muse Code terminal agent beta; Meta Model API access expands globally.

Muse Spark 1.2

Coding-focused upgrade with more coding compute and environment diversity; available via Meta Model API with expanded global access.

Muse Code beta

Terminal coding agent powered by Muse Spark 1.2 with persistent background subagents and a replay-exact event-log runtime.

Muse Spark 1.2 sibling context

Closed Muse Spark 1.2 / Muse Code beta ship days earlier as the paid Meta Model API path; Glimmer is the open local counterpart.

Qdrant 1.19 Turbo4 + memory tiers

Turbo4 4-bit TurboQuant storage datatype, unified pinned/cached/cold memory tiers, per-tenant IDF for sparse search, and keyword prefix filters.

Weaviate 1.39.0

Namespaces, 4-bit rotational quantization (RQ4), hybrid MMR, dedicated Search REST API, gRPC-web endpoint, and alter-schema improvements.

Qwen3.8-Max API

2.4T / 95B-active Max-class model on QwenCloud (qwen3.8-max) for coding and long-horizon work.

Google Workspace MCP plugins

Marketplace plugins connect agents to Gmail, Google Drive, and Calendar without leaving the IDE.

V4-Flash-0731 agent upgrade

Official Flash release with stronger agentic post-training; MIT weights on Hugging Face; same deepseek-v4-flash API ID.

Terra/Luna API price reductions

Luna −80% and Terra −20% API pricing; Sol unchanged. Fast mode replaces Priority Processing in the API.

Milvus 3.0.0 lake-native GA

External Collections index lake files in place (Parquet, Lance, Iceberg, Vortex); richer server-side retrieval (sparse SINDI, faceted search, schema evolution). Apache-2.0.

Grok 4.5 in GitHub Copilot

xAI’s Grok 4.5 becomes available as a Copilot model option across VS Code, CLI, and cloud agents.

Kimi K3 open weights

Full checkpoint published on Hugging Face under the Kimi K3 License.

Open weights on Hugging Face

Full 2.8T checkpoint published under the Kimi K3 License (not MIT/Apache).

Claude Opus 5 GA

Opus 5 ships as Anthropic’s everyday high-capability tier for agentic coding and 1M context at the same API rates as Opus 4.8.

Legacy chat/reasoner aliases retire

deepseek-chat and deepseek-reasoner stop after Jul 24 2026 UTC; traffic moves to V4 Flash.

Cursor Router (Auto mode)

Auto mode routes requests by task complexity with Cost, Balance, and Intelligence optimization modes.

Gemini 3.6 Flash GA

Flash workhorse ships generally available; 3.5 Pro remains in partner testing.

Gemini 3.6 Flash GA

Flash workhorse for agentic/coding/volume; Google confirms 3.5 Pro still testing with partners.

Kimi K3 launch

2.8T MoE multimodal agentic model ships on Kimi product and API.

platform.kimi.ai

OpenAI/Anthropic-compatible API for Kimi K3 and related models.

Kimi K3 API / product launch

Moonshot launches Kimi K3 via platform.kimi.ai and consumer Kimi surfaces.

KDA + Stable LatentMoE

Kimi Delta Attention and 16-of-896 expert MoE target long-context efficiency.

Native multimodal + 1M context

Text, image, and video in one open model with a 1,048,576-token window.

Muse Spark 1.1 + Meta Model API

Paid multimodal agentic model via Meta’s first public Model API preview.

Muse Spark 1.1

Major gains in agentic coding, tool/computer use, and multimodal understanding.

Meta Model API public preview

Meta’s first paid model API (OpenAI-compatible) opens Muse Spark to developers.

Thinking mode in Meta AI

Consumer Meta AI / meta.ai surfaces expose Muse Spark Thinking mode.

GPT-5.6 family (Sol / Terra / Luna)

Capability tiers make routing explicit: Sol for peak, Terra balanced, Luna cost-efficient.

gpt-5.6 API routing

API aliases route to Sol by default with Terra/Luna for cost control.

Claude Sonnet 5

Sonnet 5 ships as the speed/intelligence balance for most production Claude traffic.

Introductory pricing window

Aggressive intro rates push Sonnet as the default production Claude.

Claude Fable 5 launch

Most capable widely released Claude for long-horizon agents and deep reasoning.

Mythos 5 (Glasswing)

Limited-access peer of Fable for approved Project Glasswing customers.

claude-fable-5 API

Generally available on Claude API and major cloud partners at premium rates.

Adaptive thinking + effort

Effort parameter replaces manual thinking budgets for depth control.

1M context default

Million-token context and large max outputs for long agent threads.

Agent Builder and Evals platform shutdown window

OpenAI is winding down the hosted Agent Builder and Evals platform (distinct from the open-source openai/evals GitHub repo). Existing evals become read-only on 2026-10-31; the Evals dashboard/API and Agent Builder are scheduled to shut down on 2026-11-30. ChatKit remains available. Migration paths: Agents SDK code export or ChatGPT Workspace Agents.

MCP and agent tooling deepen

Tool use, structured output, and MCP-style workflows become first-class for agents.

1M-token context class

Document-heavy analysis and long agent threads become practical.

Long-context Sonnet

1M-token-class context lands on the production tier, not only Opus.

Multimodal + long context

Text, image, and audio understanding with million-token-class context for research agents.

DeepSeek V4 Preview

V4-Pro and V4-Flash bring 1M context and thinking modes to the DeepSeek API.

DeepSeek V4 Preview

V4-Pro (1.6T/49B) and V4-Flash (284B/13B) ship with 1M context and thinking modes.

deepseek-v4-pro / deepseek-v4-flash

Official API IDs replace assuming historic chat/reasoner names map to V3/R1 weights.

Muse Spark first release

Meta Superintelligence Labs ships its first closed multimodal Muse model.

Sarvam 30B and 105B open-sourced

From-scratch reasoning MoEs trained in India. 30B (2.4B active) powers Samvaad; 105B (10.3B active, 128K context) powers Indus. Apache-2.0 weights on Hugging Face.

Gemini 3.1 Pro Preview

Current Pro-class pin (gemini-3.1-pro-preview) for hard multimodal reasoning while 3.5 Pro remains in partner testing.

MCP support expands

Model Context Protocol tooling broadens Claude’s agent integration surface.

Claude Haiku 4.5 era

Fast Claude tier with stronger reasoning at economical rates.

GPT-5 family

Next major foundation-model generation across chat and API.

Sarvam-M

24B Apache-2.0 multilingual model, post-trained from Mistral Small, with an OpenAI-compatible API.

Claude 4 updates

Next-generation Opus and Sonnet strengthen coding, agents, and long-document workflows.

Llama 4

Next open multimodal generation for self-hosted stacks.

Llama 4 family

Open-weight multimodal Scout / Maverick-class models for research and commercial use.

MoE variants

Mixture-of-Experts variants expand efficiency for open multimodal serving.

Qwen3 generation

Multilingual open models spanning chat, reasoning modes, coding, and multimodal variants.

Thinking / reasoning modes

Explicit reasoning modes make Qwen competitive on hard multi-step tasks.

Gemini 2.5 Pro

Stronger reasoning and long-context performance for Vertex AI and AI Studio.

Gemini 2.5 Pro class (legacy)

Earlier Pro-class Gemini for advanced reasoning and native multimodal workloads.

Reasoning benchmark wave

GPQA / LiveBench-style comparisons put R1 on the open-reasoning map.

Industry cost/performance shock

R1 coverage forces hyperscalers and labs to revisit pricing and open-weight strategy.

DeepSeek-R1

Reasoning model resets expectations for open-weight chain-of-thought performance.

DeepSeek-R1 technical report

RL-trained reasoning model with public traces reshapes open-model expectations.

Open weights release

Weights and distillations enable self-hosted reasoning baselines worldwide.

DeepSeek API availability

Hosted API makes R1 accessible without local GPU fleets.

Production high-volume default

Becomes the default Claude choice when latency and unit economics dominate.

DeepSeek-V3

MoE open-weight model delivers frontier-competitive quality at lower cost.

V3 generalist context

Strong MoE generalist precedes R1 and pairs with it in many stacks.

Computer-use style workflows

UI and tool-driven agent patterns push Claude into operational automation.

Enterprise adoption

Privacy modes, team controls, and org rollouts make Cursor a standard engineering tool.

Agent mode expansion

Multi-step agent workflows push Copilot closer to AI-native IDE competitors.

o1 reasoning models

Models optimized for multi-step reasoning before answering.

Reasoning-model era (o1 lineage)

OpenAI popularizes deliberate reasoning traces that later feed GPT-5 thinking modes.

Cortex AI expansion

Deepens SQL-native and app patterns for LLMs on enterprise data.

Qwen2.5 momentum

Strong coding and size-ladder releases build the Qwen open ecosystem.

Studio / debugging experience

Visual debugging and graph inspection improve day-to-day agent development.

Production hardening

Metrics, continuous batching, and model coverage deepen for real traffic.

Production RAG patterns

Indexing strategies, reranking, and eval loops become standard guidance.

Claude 3.5 Sonnet era

Sonnet becomes the breakout Claude for coding and everyday agent work.

Mosaic AI expansion

Deepens train/serve/evaluate for enterprise RAG and agent workloads.

Tool use on Haiku

Reliable tool calling makes Haiku practical for high-volume agent workers.

Long-context Gemini push

Million-token-class context becomes a Gemini product differentiator.

vLLM serving maturity

Production serving recipes for Qwen solidify across open inference stacks.

LangGraph Platform momentum

Managed deployment and ops story expands beyond the open-source library.

Inference / embedding services

Vector DB expands toward integrated embedding and inference workflows.

Multi-model backend era

Teams routinely switch among GPT, Claude, and other models inside the same IDE.

Human-in-the-loop patterns

Interrupt/resume flows make approval steps practical in production agents.

Distributed production deployments

Clustering and scale-out patterns mature for larger self-hosted estates.

Arctic models

Signals Snowflake as both a data platform and a model contributor.

Llama 3

Major quality jump for open-weight chat and coding models.

Llama 3 quality jump

Major open-weight quality leap that sets up the Llama 4 generation.

LangSmith production focus

Observability, evaluation, and deployment tighten the path from prototype to production.

LoRA / multi-adapter serving

Serving many fine-tuned adapters efficiently becomes a production requirement.

DBRX open model

Signals Databricks as both a platform and a model contributor in open LLM space.

Blackwell architecture

Next-generation accelerated computing platform for training and inference at AI-factory scale.

NVIDIA NIM

Microservices for deploying optimized inference in enterprise AI factories.

Claude 3 family

Opus, Sonnet, and Haiku establish a clear quality–latency–cost ladder for production.

Claude 3 Opus launches

Establishes Opus as Anthropic’s peak-capability brand in the Claude family.

Claude 3 Sonnet launches

Defines the mid-tier Claude brand for balanced cost and capability.

Claude 3 Haiku launches

Defines Haiku as Anthropic’s speed and cost brand in the Claude family.

Vision on Haiku

Image understanding lands on the fast tier for multimodal pipelines.

Checkpoints & persistence mature

Durable state, retries, and long-running workflows become first-class.

Workflows / agents expansion

Framework grows from pure RAG into broader workflow and agent patterns.

Mistral Large

Flagship commercial model strengthens La Plateforme for enterprise APIs.

Le Chat

Consumer/business chat surface for Mistral models.

Gemini 1.5 era

Native multimodal + long context establishes Gemini’s modern identity.

Generative / RAG features

Closer integration with generative search patterns for end-to-end RAG apps.

Copilot Enterprise era

Org knowledge, admin controls, and enterprise packaging deepen GitHub-native AI.

LangGraph open release

Graph-based runtime for durable, stateful LangChain agents.

LangGraph introduced

Stateful graph runtime complements LangChain’s higher-level abstractions.

DeepSeek API platform

OpenAI-compatible API makes DeepSeek easy to try in existing clients.

Hugging Face distribution

Hub becomes the default discovery path for Qwen weights and demos.

Serverless / usage-based era

Packaging shifts toward more elastic, usage-based vector infrastructure.

Agentic coding features expand

Multi-file edits and agent-style workflows deepen IDE autonomy.

Ecosystem default for open serving

Becomes a standard choice next to TGI and TensorRT-LLM in many stacks.

Phi small models

Shows competitive quality from small, efficient models for edge and cost-sensitive use.

Mixtral 8x7B

Sparse MoE open model becomes a default efficient alternative to dense LLMs.

Gemini 1.0

Natively multimodal foundation model family launches across consumer and cloud surfaces.

Gemini family launches

Google’s unified multimodal foundation-model brand debuts.

Seed and Series A (~$41M)

About $41 million raised, with Lightspeed, Peak XV Partners, and Khosla Ventures among the backers named by Sarvam.

Assistants API

Platform primitives for tool use, retrieval, and agent-style apps.

Snowflake Cortex

Brings LLM functions and AI services into the governed Data Cloud.

Hybrid search capabilities

Sparse + dense retrieval patterns expand beyond pure vector similarity.

Mistral 7B

Small open model punches above its size and accelerates European open LLM adoption.

Multi-file editing / Composer

Repo-aware edits across files shift Cursor from autocomplete to agentic coding.

OpenAI-compatible serving

Drop-in API compatibility accelerates migration from hosted APIs to self-host.

Sarvam AI founded

Vivek Raghavan and Pratyush Kumar start a Bengaluru lab focused on AI for India.

Expression Language / LCEL era

Composable runnables become the preferred way to build chains and pipelines.

Evaluation & observability focus

RAG quality tooling expands beyond connectors and indexes.

Llama 2 open release

Openly available weights for commercial and research use.

Llama 2 commercial license

Open weights become broadly usable for commercial products under Meta’s terms.

Claude 2

Longer context and stronger conversational performance for enterprise assistants.

DeepSeek founded

Hangzhou lab focuses on efficient, high-capability open models.

MosaicML acquisition

Brings training and generative AI tooling into Mosaic AI on the Lakehouse.

Hybrid / metadata retrieval focus

Sparse-dense and metadata filtering become table stakes for enterprise RAG.

PagedAttention / vLLM rise

Open serving engine makes high-throughput LLM inference widely accessible.

GPU indexing momentum

GPU-accelerated indexing deepens Milvus’s large-scale search positioning.

AI infrastructure surge

GPUs become the default substrate for training and inference at global scale.

Hybrid search prominence

Keyword + vector retrieval becomes a primary reason teams evaluate Weaviate.

Google DeepMind formed

Google Brain and DeepMind combine—research, models, and product under one AI organization.

Mistral AI founded

Paris-based lab founded by former DeepMind / Meta researchers.

Google Brain × DeepMind

Research and model product organizations consolidate under one AI org.

Data connector ecosystem

Ingestion from docs, APIs, and SaaS sources becomes a core product strength.

Copilot Chat

Chat moves Copilot beyond autocomplete into conversational coding help.

GPT-4 introduced

Multimodal foundation model with stronger reasoning and coding.

Integration explosion

Model, vector-store, and tool connectors make LangChain the glue layer for LLM stacks.

Payload filtering strength

Rich metadata filters become a defining reason teams pick Qdrant for RAG.

Cursor IDE gains traction

AI-native fork of VS Code becomes a default coding environment for many teams.

LLaMA to researchers

Research weights that sparked the open-model wave.

LLaMA research weights

Research release that sparks the modern open-model wave.

Expanded OpenAI investment

Deepens Azure OpenAI as the enterprise channel for ChatGPT-class models.

Hub as open-model hub

Models, datasets, and Spaces concentrate open AI collaboration and distribution.

Moonshot AI founded

Beijing lab builds the Kimi assistant and frontier MoE models.

ChatGPT released

Consumer chat product that brought LLMs into mainstream use.

LlamaIndex (GPT Index) adoption

Becomes a leading open framework for connecting LLMs to private data.

Inference Endpoints

Managed production inference for Hub models without self-hosting ops.

LangChain gains adoption

Becomes the default open framework for composing LLM apps.

GitHub Copilot GA

Generally available as a paid coding assistant for individuals and teams.

Pod-based scale era

Production customers standardize on managed indexes for semantic search and RAG.

Qdrant Cloud

Managed option alongside self-hosted deployments.

PaLM

Pathways Language Model showcases large-scale Google LLM research before Gemini.

Hopper architecture (H100)

Defines the GPU generation that trains and serves most frontier models of the mid-2020s.

Weaviate Cloud

Managed offering expands beyond self-hosted deployments.

Zilliz Cloud managed offering

Managed Milvus becomes the enterprise path without self-host ops.

Spaces

Turns the Hub into a place to demo and share interactive ML apps.

GitHub Copilot

Coding assistant brings LLM copilots into everyday developer workflows.

GitHub Copilot technical preview

Inline AI completions bring LLM coding assistants into mainstream IDEs.

Modules / vectorizers

Pluggable modules make embedding and enrichment part of the DB workflow.

CNCF sandbox / cloud-native path

Milvus joins the cloud-native ecosystem path for large-scale vector infra.

Pinecone product traction

Managed vector search becomes mainstream for production RAG.

Anthropic founded

Safety-focused lab founded by former OpenAI researchers; Constitutional AI becomes a defining idea.

Qdrant open-source growth

Rust vector engine gains adoption for filtered similarity search.

Milvus 2.x distributed architecture

Cloud-native redesign enables billion-scale indexes and separation of storage/compute.

Snowflake IPO

Public listing accelerates enterprise Data Cloud adoption.

OpenAI API launched

Developers gain API access to GPT models for production apps.

Datasets library

Standardizes how the community loads, shares, and versions training data.

Lakehouse positioning

Unifies data warehousing and data lakes as the substrate for analytics and AI.

Milvus open-sourced

Zilliz releases Milvus as an open vector database for similarity search.

Microsoft strategic partnership

Microsoft invests $1B and becomes a primary cloud partner.

OpenAI partnership begins

Multi-year investment and Azure exclusivity set up GPT distribution.

Weaviate open-source vector DB

Early open vector database with a modular architecture.

Transformers library

Becomes the default open toolkit for pretrained NLP models.

MLflow open-sourced

Becomes the default open toolkit for experiment tracking and model registry.

Transformer architecture published

“Attention Is All You Need” (Google Research) becomes the foundation of modern LLMs.

TensorRT era

Inference optimization stack becomes central to production deep learning serving.

PyTorch released

Becomes the dominant research framework and later the industry standard for training modern neural nets.

AlphaGo vs Lee Sedol

Landmark demonstration of deep reinforcement learning on a grandmaster-level game.

Hugging Face founded

Started as a chat app company; pivoted into open ML tooling.

OpenAI established

Founded as a nonprofit AI research lab in San Francisco.

Databricks founded

Created by the creators of Apache Spark to commercialize big data and ML platforms.

Snowflake founded

Cloud data platform company founded to separate storage and compute.

DeepMind founded

London research lab focused on general-purpose AI systems.

Azure cloud era

Azure becomes the cloud substrate for later AI services and OpenAI hosting.

CUDA introduced

GPU programming platform that later underpins modern AI training and inference.

Microsoft founded

Started as a microcomputer software company in Albuquerque.