Last posted 15 Oct 2026, 09:07 IST: 121 changelogs checked, 14 new entries.

OpenTelemetry export (beta) for Messages API and Managed Agents

Trace Messages API requests and Managed Agents sessions in your own observability backend; SDK spans in 7 languages, API spans need an OpenTelemetry destination set in the Console.

Claude Code 2.1.299

Idle compaction for conversations over 200K tokens before cache expiry; MCP allow rules now match only the named server (mcp__a no longer matches a__b); MCP 2026-07-28 negotiated by default for claude.ai connectors.

Gemini 3.8 Flash GA in the UK with in-region processing

Gemini 3.8 Flash is generally available in europe-west2 (London) with in-region data residency and ML processing, on by default in the web app there.

HydraFusion Max and a bigger HydraFusion model pool

Copilot's HydraFusion now draws on Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna; a new HydraFusion Max option targets harder tasks.

OpenCode v2.0.29

Adds remote device pairing via OpenTunnel (QR code or link), provider account switching, and automatic continuation when a response hits the output-token limit.

Zed 1.24.2

Terminal tabs can be dragged into the Agent Panel, custom Amazon Bedrock models get tool use, images and thinking, and long agent threads use less memory and CPU.

Notion search filter_properties; MCP backlinks and groups tools

Search accepts filter_properties to return only chosen page properties; Notion MCP adds notion-list-page-backlinks and notion-get-groups tools.

Workflow Builder connects to Microsoft Teams, Notion and Salesforce

Gemini Enterprise Workflow Builder workflows can now search data in and automate actions across Microsoft Teams, Notion and Salesforce.

Antigravity in the Gemini Mac app; Customize menu with 40+ apps

Antigravity coding agents can be started from chat in the Gemini macOS app; a new Customize menu adds skills and connects 40+ apps, rolling out on web, mobile and desktop.

Agents SDK persistent browser tool; API MCP server sends Worker issue events

Agents SDK adds browserTool, a Browser Run browser that keeps tabs, cookies and logins across turns; the Cloudflare API MCP server now pushes signed Worker-issue events to subscribed clients.

Post-training on Together AI

New Post-training product lets you write your own training loop on Together GPUs with SFT, DPO and GRPO methods, billed per GPU-minute.

MAI-Image-2.6 and MAI-Image-2.6-Flash on AI Gateway

Microsoft's MAI-Image-2.6 and 2.6-Flash image generation and editing models are available through Vercel AI Gateway with one key and unified billing.

Kiro IDE 1.3.0

Local command sandbox limiting what agent-run shell commands can access (block/restrict/open network), /workflow-run to start saved Workflows from chat, on-demand MCP tool loading.

Kiro CLI 2.31.0 (and 2 earlier builds)

Local command sandbox and local autonomous mode for all V3 users, on-demand MCP tools; 2.30.0 starts the gradual V3 rollout with model-fallback notices.

Origin App Marketplace

Cursor's Origin integrates with 20+ apps for deployment, CI, code review, security, monitoring and project management; the marketplace accepts public submissions.

Copilot managed-settings enforcement and telemetry user attribution

Admins can require Copilot clients to fetch managed settings at startup, with minimum client versions enforced for governance keys starting with sandbox; client telemetry can now attribute activity to users (12 Oct).

Lovable WebMCP tools for browser agents, image editing, Loops connector

Apps can expose WebMCP tools that browser agents run in the user's tab; image editing in the Files tab; Loops connector; fixed list of supported Postgres extensions; free-allowance coverage in usage.

OpenCode v2.0.28

Per-project auto-accept for permissions, native Windows ARM64 builds, official WinGet packages, optional Figma plugin with MCP, and 15-minute streamed-chunk timeout for reasoning models.

Skills in Workspace delayed; Gemini app skills roll out to mid-November

Gemini app skills began a gradual rollout on 13 Oct, finishing by mid-November; skills inside Workspace apps are delayed until further notice.

Notion 3.8: Context for your agents

MCP-compatible agents can read and update Notion; Notion agents use context from 150+ MCP tools; Skills API syncs skills with ChatGPT and Claude; GPT-6.1 Sol, Opus 5.5 and Sonnet 5.5 added; admin MCP controls.

Notion MCP can delete pages and databases

Two new Notion MCP tools move up to 20 pages or databases to Trash after user confirmation; they can be restored.

Kling 2.5 Turbo Pro, Kling 3.0 and Kling O3 on Runway Dev

Kling text, image and video generation models (including 4K variants) are available in the Runway API, billed per second.

Langfuse redesigned session view; custom System One decision models

Sessions show messages across traces as one conversation with a navigation sidebar; the System One adapter accepts a custom base URL for decision-model evaluators.

Instant Voice Cloning in the Python and TypeScript SDKs

Instant Voice Cloning (beta) is now exposed in the sarvamai SDKs with one-shot, HTTP streaming and WebSocket calls; needs Python 0.1.37+ or TypeScript 1.1.12+.

ZCode 3.15.1 (and 1 earlier build)

Adds MCP App UI plugin support, per-sub-agent model and tuning options in workflows, and a referral campaign, plus bug fixes.

Agents SDK adds web_fetch and web_search tools; OpenCode 2.0 harness

Agents SDK gains a web_fetch tool (webpages, PDFs and Office docs to Markdown) and a web_search tool, plus an experimental harness to run OpenCode 2.0 in a Durable Object.

Contributor permission for access tokens

Fine-grained and service-account tokens can be allowed to create repositories, with write access only to repos they created.

Ollama v0.40.4 (and 1 earlier build)

Adds decision-model support in the CLI (ollama run --ask), Gemma 4 prefill speed-ups on NVIDIA via MLX runner, multimodal embeddings via llama-server and speculative-decoding fixes.

Devin Desktop v3.10.67

Supports MCP servers on the 2026-07-28 spec, lets enterprise admins lock extensions to one marketplace URL, new folder/git @-mentions and private-preview remote MCPs from Devin Cloud.

Antigravity 2.23.0

Permissions preset picker in the message box, full-screen image/video viewer, .zip and folder attachments, and a 'Fix with Agent' option when an MCP server fails to start.

OpenCode v2.0.27

Built-in worktree tools in Code Mode, confirmation before 'opencode pair --remote' exposes a server, installable web app, and Vertex AI routes for Mistral models.

Gemini next steps on Google Meet post-meeting screen

After a meeting with 'Take notes for me' on, Gemini suggests next steps: draft follow-up emails in Gmail or export action items to Google Tasks.

TTS audio effects, turn-prediction STT; seed field deprecated

TTS and Text to Dialogue accept audio_effects at no extra cost; Realtime STT adds turn_prediction commits; call queueing on by default for new agents; TTS seed and KB auto_discover deprecated.

LiteLLM v1.105.0

Day-one Claude Sonnet 5.5 and GPT-6.1 Sol support, Decisions API (/v1/decisions, /v1/systemone) across eight providers, new Prism, Sail and Cortecs providers, and litellm.agent(); check breaking changes before upgrading.

OpenCode v1.19.0

Adds 'opencode upgrade --major' and a /update command to move to OpenCode 2 (updates no longer auto-install), allows OpenAI's ultrafast service tier, and hardens PTY connection checks.

Session affinity for Chat Completion V2

Optional session-affinity header on gemma4 at POST /v2/chat/completions lets multi-turn chats reuse cached prompt context for lower latency.

Also announced by Sarvam docs

Cloudflare API MCP server serves Cloudflare skills

The Cloudflare API MCP server now provides Cloudflare skills to MCP clients that support the Skills over MCP extension.

New controls and chat improvements in Copilot for JetBrains

Copilot for JetBrains adds enterprise-managed default models, a Fix action for diagnostics and an MCP auto-startup setting, and drops support for JetBrains IDE 2025.1.

Dynamic workflows for Managed Agents (beta)

Managed Agents can write and run multi-agent workflows in the background, enabled via the multiagent_20261001 config.

Claude Code 2.1.296 (and 2 earlier builds)

Hook onFailure block, higher MCP tool-description limits, MCP 2026-07-28 negotiated by default for claude.ai connectors, subagent autoCompactWindow, Read allow_large option.

Composer predictions in Codex (beta)

Codex suggests your next message after a response; free during beta for Pro users aged 18+, and predictions don't count toward usage limits.

Decision models on the Hub

Decision models get their own tag, icon and Models-page filter; GGUF versions are tagged automatically from llama.cpp metadata.

AI Gateway adds Liquid d1 and Step 5 Preview; agents can buy domains via CLI

Liquid AI d1 and StepFun Step 5 Preview available on AI Gateway; Vercel CLI lets agents purchase domains; Sandbox creation now returns a ready sandbox.

Kiro CLI 2.29.0

Local V3 sessions can remember useful context and reuse it in later chats, with read, write or off controls.

Lovable: workspace publishing domain, cross-chat tasks, 10 new connectors

Enterprise workspaces can publish all apps under their own domain; chats can reference or task other chats; new connectors incl. Zendesk, Neon, Oracle, IBM Db2, Metabase, Loom; invite-only publishing on Free/Pro.

Replit 9 Oct release

Replit Desktop for Windows in private preview (with Microsoft and NVIDIA), sandboxed local builds; send tasks to existing projects from one chat; TikTok Ads via MCP.

Workspace Studio adds logic and Drive search steps

Workspace Studio automations gain nested conditionals, Gmail reply skipping, Meet calendar outputs and a Drive 'Search for a Doc' step.

Notion API: silent page writes and schema-update fixes

Page write endpoints accept a silent notifications mode; data-source schema updates fix option colours and stop clearing omitted descriptions (behaviour change).

n8n 2.42.6 (and 4 earlier builds)

Blocks creating/updating workflows that use deprecated nodes; adds Nano Banana 2.1 and 2 Lite to the Gemini node; Kimi K3 fix in the self-hosted Assistant (2.43.x pre-releases).

Compliance API covers unified Claude experience chats (beta)

For Claude Enterprise, Compliance API chat endpoints now include unified Claude experience chats; existing access keys work. Models API also reports server-tool support (6 Oct).

Compliance API can download Claude Docs as Word files (beta)

Claude Enterprise admins can now pull Claude Docs documents out through the Compliance API as Word (.docx) files, in beta.

Claude Code 2.1.294 (and 7 earlier builds)

Since 2.1.287: Claude Mods for plugins, Haiku 5.5 as default Haiku, local stdio MCP servers default to protocol 2026-07-28, plugin install --marketplace, claude purge.

Gemini 3.8 Flash GA in Singapore with in-region processing

Gemini Enterprise offers Gemini 3.8 Flash in Singapore with in-region data residency and ML processing.

Ultrafast mode for GPT-6.1 Sol

Responses API adds an ultrafast service tier for gpt-6.1-sol to cut time between output tokens, for all API users subject to rate limits; no price stated.

Also announced by OpenAI release notes, OpenAI Codex

Faster steering in Codex; plugin access by role for Enterprise/Edu

Codex in the ChatGPT desktop app reacts sooner to follow-ups during a run; Enterprise/Edu admins with RBAC can set plugin access per role.

LM Studio 0.4.26

Adds /v1/systemone and OpenAI-compatible /v1/decisions endpoints, Bionic coexistence, CUDA 13 on Windows ARM and llama.cpp runtime 2.50.0.

Kiro IDE 1.2.56 (and 1.2.37)

Adds a Workflows feedback link, improves workflow reliability, spec-only commands and hook matching, and fixes chat, terminal and source-control issues.

Copilot code review: organisation billing options and controls

Org owners can bill reviews requested by licensed members to the organisation's cost centre; admins can limit review requests to users with org or enterprise licences.

Antigravity 2.22.0 (and 1 earlier build)

Search across conversation contents, Scheduled Tasks renamed Automations, grouped agent actions, sidebar drag-and-drop and UI extensions in side pane.

AQuA open-sourced; ML Drift for on-device GPU inference

Google open-sources AQuA, a reference agent that diagnoses quality failures of production agents on Google Cloud, and introduces ML Drift for edge GPU inference.

Midjourney testing Thinking Mode; alpha changelog 7 Oct

Midjourney posted 'Testing Thinking Mode' and an alpha changelog; the index page shows titles only.

Suno Studio timing, clip stretching and fade curves

Studio gets more accurate downbeat detection, shift-drag clip time-stretching, curved fades and pop-out plug-in windows.

Browser-use and computer-use SDK classes (beta)

Python and TypeScript SDKs add classes that run the browser/computer-use tool loop, URL and file policies and approval callbacks.

Codex CLI 0.161.0 (and 1 earlier build)

GPT-6.1 Sol becomes the default model (incl. Bedrock catalog); /mcp login from a session; choose mic/speaker for voice; Bedrock multi-agent V2 support.

chat-latest now tracks the newest ChatGPT model

The chat-latest snapshot follows the newest model in ChatGPT for paid business plans; OpenAI recommends GPT-6 models for production.

Serverless rate limits doubled for Small models

Small serverless models (<600B params) now allow 216M total prompt TPM, 54M uncached and 2.16M generated TPM.

Copilot local sandboxing GA; local Ollama models in CLI; AI secret-detection model

Local sandboxing GA in Copilot CLI, app and VS Code at no extra cost; CLI /model discovers local Ollama models; purpose-built model for AI secret alerts.

Antigravity CLI 1.3.1 (and 2 earlier builds)

Default verbosity changes to medium in 1.3.0; /diff gets cursor and file navigation; Windows command sandbox runs without admin rights (1.2.17).

Lovable: workspace credit limits, Drizzle migrations, new connectors

Admins can set credit limits and alerts; Cloud DB schema changes saved as Drizzle migrations; Klaviyo, Excalidraw+, beehiiv, Ramp connectors; Lyria 3 music models.

Warp 2026.10.07

Kiro CLI joins the third-party coding-agent toolbar; unattended CLI agent runs end as blocked instead of hanging; Windows computer-use agents can record video.

Zed 1.23.2

Adds agent idle-thread retention setting, automatic retries when Google models are overloaded, and better ACP plan updates.

Developer Knowledge API and MCP server (public preview)

Public preview of an API and MCP server that lets AI assistants search current docs across Firebase, Google Cloud, Android and other Google products.

Langfuse: API key expiry, decision-model evaluators, usage by project

API keys can have expiry dates; observations can be scored with gpt-6-luna decision models; new organisation usage chart by project with CSV export.

Admin control over Antigravity Boost and Teamwork

Admins can control user access to Antigravity Boost (GA) and Teamwork (Preview) from Gemini Enterprise.

Copilot Search chat, Regenerate/Switch Model, PowerPoint custom skills

Copilot Search gains in-results chat, responses get Try Again and Switch Model, PowerPoint gets custom skills, and Cowork adoption metrics land in Viva Insights.

Audio uploads in ChatGPT

Paid subscribers and workspaces can upload audio files for transcripts, summaries and questions; availability varies by settings, region and model.

AI Gateway: confidence-based decision fallbacks; Glyph Cluster

AI Gateway can escalate a decision to a fallback model when confidence fails a set condition; stealth model Glyph Cluster is free to call for now.

Kimi K3 India cross-Region inference on Bedrock

Kimi K3 is available through the in.moonshotai.kimi-k3 inference profile in Asia Pacific (Mumbai) and Asia Pacific (Hyderabad).

Remote control for local agents

Users can view and reply to agents running on their computer from the Cursor iOS app.

Update IDE to restore agent activity in Copilot usage metrics

IDEs using the Copilot SDK for agent mode left agent activity out of usage metrics; fixed in VS Code 1.139.0+, other IDEs by November 2026.

Gemini CLI v0.63.0

Enables autonomous plan execution in non-interactive mode and bounds tool output and memory in long-running agent loops.

OpenCode v1.18.35

Adds JSON and Markdown formats for agent-readable stats; xAI tool results now carry supported images.

Models API adds model line and thinking-disable flag

GET /v1/models now returns each model's line (e.g. opus) from 1 Oct, and from 5 Oct reports whether a model accepts thinking: {type: "disabled"}.

Self-serve HIPAA BAA for API organisations

Eligible API organisation admins can accept the standard Business Associate Agreement and turn on HIPAA support from Organization settings.

Also announced by OpenAI release notes

Automations, inline visualisations and new models

Perplexity adds Automations for ongoing work, interactive visualisations in threads, GPT-6.1 Sol, Claude Opus 5.5, MiniMax H3, Seedance 2.5 and Amex business Skills.

P(doom) field on user profiles

Users can add an estimated probability of AI existential catastrophe to their profile, feeding an anonymised community survey.

llama.cpp v0.6.0 (and builds to b11435)

v0.6.0 adds GLM-5.3-Flash, Clef, Ling 3.0 VL and Nimble support, a /v1/systemone decision-model API, llama_batch_ext and a web UI overhaul.

Kiro CLI 2.28.0 (and 1 earlier build)

Classic users can switch to 3.0 from an interactive session, and V3 hooks can match tools by category.

Stock charts in Google Sheets

Sheets adds High-Low-Close and Open-High-Low-Close stock charts, which now survive Excel imports.

Native Markdown (.md) files in Drive and Docs

Open, edit and co-author .md files in Google Docs without converting, with rendered previews in Drive; rolling out from 5 Oct to all Workspace and personal accounts.

Albums

Creators can publish a collection of their work as an album, which notifies their followers.

New and improved Extract (Doc Agents)

Extract is more accurate on Indic, English and mixed-language documents, with lower latency, longer documents, per-value citations and higher rate limits.

Federated query mode for data connectors (Preview)

Gemini Enterprise can query AlloyDB, BigQuery, Cloud SQL and Spanner data in place without ingestion; Knowledge Catalog integration.

Finances expands to Free and Go users (US)

Financial account connection for spending insights and investment tracking now available to Free and Go users in the US on web, iOS and Android.

Also announced by OpenAI release notes

Web Search API (beta)

AI agents can search the web through Ceramic.ai, Exa or Linkup with zero data retention, billed through AI Gateway credits.

Pi Durable harness on the Agents SDK (beta)

Run Earendil's Pi framework as a long-running agent harness with durable persistence on Cloudflare.

Higher serverless rate limits for Small models

Adaptive rate-limit ceilings raised for serverless models under 600B parameters.

Provisioned throughput adds GLM-5.3 and DeepSeek-V4.1-Flash; dynamic rate limits retired

GLM-5.3, GLM-5.3-Flash and DeepSeek-V4.1-Flash available on provisioned throughput; serverless no longer applies traffic-scaled dynamic rate limits.

Copilot code review API support; Balanced default effort

Request Copilot code review through REST and GraphQL with configurable effort; Balanced is now the default level.

Lovable: EU AI data residency, API-key IP allowlist, GPT Live voice

Enterprise EU inference option, IP allowlisting for workspace API keys, announcement banners, GPT Live voice in TanStack Start apps, Lovable MCP server with Meta Muse.

Replit 2 Oct release

Jev model via Replit integrations, GPT-6.1 Sol and Claude Sonnet 5.5 added, Settings now a full page, workspace-level configuration controls.

Change Nova-3 keyterms mid-stream

Replace vocabulary during a live /v1/listen session with Configure messages, no reconnect; Flux STT numerals can also be toggled mid-stream.

Scan notes and documents with ChatGPT camera

The iOS camera can scan multi-page documents and compile them into a single PDF automatically.

Also announced by OpenAI release notes

New ways to shop in ChatGPT

Virtual try-on for clothing and accessories, plus saving products to favourites folders, on web and mobile.

Also announced by OpenAI release notes

Voice Cloning streaming

Voice cloning now streams audio over a single HTTP stream or a persistent WebSocket that accepts incremental text.

Also announced by Sarvam docs

Workers OAuth Provider v1 with MCP 2026-07-28 support

Separate authorization and resource servers, MCP 2026-07-28 compatibility and service-binding token validation.

Code sandbox SDK and CLI

together-sandbox SDK (Python, TypeScript) and CLI run commands in isolated Docker environments using Together API keys; replaces the legacy CodeSandbox integration.

Kiro CLI 2.27.0 (and 2 earlier builds)

V3 gets multi-step Workflows, workflow delegation controls, live steering context and saved prompts as slash commands; Classic interface shows deprecation guidance.

Dynamic workflows in Copilot CLI, app and SDK

Define orchestrated multi-agent processes in code.

Copilot in VS Code: September 2026 releases

VS Code 1.136 to 1.140 add agent automations, agent merge for pull requests and better session management.

Copilot computer use in public preview

Copilot CLI and the Copilot app can control desktop applications on macOS and Windows.

Gemini student hub for Workspace for Education

Central place for study tools such as notebooks and flashcards, now available to Workspace for Education users.

Meta Model API: Muse Code GA, Muse Image API, Muse Glimmer (undated)

Undated items on the hub: Muse Code leaves beta with new plans, Muse Image on the API at $0.01 per image, and Muse Glimmer open-weight model for local agents.

AgentCore Memory: time-to-live for long-term memory records

Long-term memory records can be set to expire (minimum 31 days); setting or removing a TTL needs the new UpdateMemoryRecordExpirySpecification IAM permission.

AgentCore Gateway: private CA support for targets

TLS certificates from private certificate authorities on MCP server, OpenAPI and HTTP proxy targets, allowing direct connections to VPC endpoints.

Admin API organisation methods leave beta

In the latest SDKs and ant CLI 1.37.0, organisation Admin API methods move to client.organization; the beta namespace stays available.

Claude Code 2.1.286 (and 5 earlier builds)

Sonnet 5.5 becomes default Sonnet (2.1.284); claude --desktop and install-time plugin config (2.1.285); deniedModels managed setting (2.1.283); Bedrock assume_role and MCP URL elicitation (2.1.281).

Kiro IDE 1.2.4

Background multi-agent workflows, safer handling of untrusted workspace configs and enterprise sign-in controls.

HydraFusion in VS Code and the Copilot app

The research preview that orchestrates multiple models now extends to VS Code and the Copilot app.

Antigravity 2.19.1 (and CLI 1.2.10 to 1.2.13)

Message subagents directly, export rendered Markdown as PDF, conversation-only undo; CLI adds --effort and server-requested retry delays.

Zed 1.22.0

BYOK for GPT-6.1 Sol, per-subagent model selection, automatic subagent compaction and a unified/split diff toggle.

Skills in the Gemini app and Workspace (successor to Gems)

Reusable custom instructions called Skills arrive in Gemini and Workspace, replacing Gems.

Also announced by Gemini app

Comments and suggestions in Docs, Sheets and Slides APIs

Developers can programmatically manage comments and suggested edits across the Docs, Sheets and Slides APIs.

SIEM event for meeting-notes consent setting

The SIEM/DLP stream emits workspace.settings.meeting_notes_consent_setting_updated when admins change consent for AI meeting-note transcription.

Flux TTS inline pause and pronunciation controls

Inline pause markers (500 to 3000 ms) and IPA pronunciation controls in batch and streaming modes.

Codex CLI 0.159.2 (and 3 earlier builds)

GPT-6.1 Sol default model incl. Bedrock catalog (0.159.1); opt-in instant interrupt (0.159.0); MCP OAuth client secrets and secure WebSockets (0.158.0).

Sign in with ChatGPT and use your plan allowance in other apps

Single sign-on for participating products, plus a limited preview letting Plus/Pro users authorise apps to draw on their ChatGPT allowance.

Also announced by Devin Desktop

'dot' always-on agents, Meetings plugin and event-triggered automations

Always-on agents on GPT-6 Astra for eligible Pro/Business Premium users; beta meeting notes plugin; automations can start from connected-app events.

Codex Cloud, Codex code review and Codex Security Cloud

Cloud environments for continuous Codex tasks across devices, a Code Review plugin for pull requests, and scheduled GitHub repository scanning with proposed fixes.

ChatGPT Pages, Space and Sites updates

Turn chats into collaborative Pages, organise work in ChatGPT Space (Pro/Business/Enterprise), and edit or schedule published Sites.

Computer use in Agents API; Ultrafast tier for GPT-6 Astra

Agents can now work in OpenAI-hosted browsers (site access, sign-in handling). New Ultrafast service tier cuts time between output tokens for GPT-6 Astra.

ZCode 3.14.4 (and 3 earlier builds)

Dynamic workflows orchestrating sub-agents and Office/Coding mode switch (3.14.0); adjustable workflow concurrency (3.14.3); CAPTCHA removed for free tier.

AI Gateway User Insights task analysis

AI Gateway groups conversations by task to show model overuse and possible savings.

Ollama 0.35.1 (and 3 earlier builds)

Decision models via /v1/systemone (0.35.0), up to ten web searches per response, MLX default on Apple Silicon in 0.40.0-rc0.

Fine-tuning: LoRA rank to 128, longer Qwen 27B context; batch files kept 7 days

LoRA rank up to 128 (from 64); Qwen 27B SFT context 131K and DPO 65K; batch job input, output and error files retained for 7 days.

Resource-level access controls; new data stores and M365 actions

Per-app and per-data-store IAM permissions; new OneDrive/Outlook/SharePoint/Teams actions and several new data stores in preview.

Personalised summaries of health data

Pick charts or metrics in the Health tab to get explanations based on connected health information.

Also announced by OpenAI release notes

Antigravity 2.18.1 (and 4 earlier builds)

Plugin install/management (2.18.1), /plan before building (2.17.0), WSL and Office attachments (2.16.0), Windows sandbox (2.15.1).

Claude Agent SDK (Python) v0.2.161 (and 4 earlier)

New verbatim_prompts option to stop untrusted text triggering file reads or commands (0.2.158); bundled CLI updated to 2.1.284.

crewAI 1.15.23

Adds native Gemini 3.8 Flash support, AMP evaluation of the last traced run in crewai eval and easier platform integration setup, plus tracing and Bedrock fixes.

LangSmith Cloud updates (21 Sep - 5 Oct)

Adds a query_runs tool to LangSmith MCP, wider OpenAI and Vertex cost tracking, Claude Opus 5.5 in Playground and Sonnet 5.5 in LangSmith Chat.

LiteLLM v1.103.0

Unified config precedence, Auto Router capability classifier, MCP allowlisting and OAuth delegation, plus 408 new model catalogue entries.

Dubbing API and Voice Cloning API

New APIs to localise audio/video while keeping each speaker's voice, and to synthesise speech in a cloned voice from a short sample.

Also announced by Sarvam docs

Build plugins for Claude

Developers can submit plugins for Claude through a developer portal with review tracking and usage analytics.

Fix for GPT-6 Sol and Luna image understanding

An image-encoding bug that degraded visual tasks, including computer use, in GPT-6 Sol and Luna has been fixed.

Also announced by OpenAI release notes

safety_identifier request field

Optional field attributes policy violations to end users rather than the API key.

Live resource usage on Jobs

Job pages show a Resources panel with live CPU, RAM, GPU and VRAM usage while a job runs.

Ollama v0.40.0

On Apple Silicon, supported models now run on MLX by default; adds gemma4, qwen3.6, qwen3.5, decision models on MLX and embeddinggemma-2. Page dates it 25 Sep.

Agentic autofix uses Copilot Memory; Slack/Teams updates

Security autofix learns from stored memory and past fixes; Copilot for Slack/Teams gets per-message model choice and duplicate-issue checks.

Replit 25 Sep release

Meta device app building, Muse integration, inline data charts, Airwallex payment testing via MCP, and enterprise access-request workflows.

Meta Connect 2026 developer recap

Meta VR Glasses unveiled with v207 SDK, three build paths for AI glasses, early access to Horizon Create and Horizon Studio, and a VR Start developer competition.

External access controls for ChatGPT Business and Enterprise

Admins can control, from the Admin Console, which sites access connected apps and which apps access ChatGPT Ads.

Hugging Face on Google Cloud Marketplace

Organisations can pay for Inference Providers, Endpoints, Spaces, Jobs and seats via Google Cloud invoices; Jobs now show live CPU/GPU usage.

TogetherLink beta; Tev1-4B-experimental

TogetherLink routes coding agents to Together models with cost tracking; Tev1-4B-experimental serverless at $0.042 input, free output.

Foundry: wider model choice, voice agents, continuous optimisation

Microsoft Foundry adds more model options, voice-agent support and automated agent performance tuning.

Kiro IDE 1.1.70

Configurable terminal timeouts for agent commands and Markdown links to workspace files.

Cloud API Gateway as an MCP server

Google Cloud API Gateway can expose REST APIs as MCP tools through OpenAPI annotations.

Edit updates, thumbnail previews

New options to edit posts and preview image thumbnails.

Cache diagnostics generally available

Cache diagnostics out of beta on the Claude API: no beta header needed and the diagnostics field is always returned. Compliance API local sessions now cover Claude in Chrome (beta, 18 Sep).

Excel Copilot change links and attribution; licence request page

Copilot responses link straight to the Excel cells it changed and Show Changes marks AI edits; admins get a Copilot licence-requests page.

Plugins in Voice conversations

Plugins now work in Voice on web, iOS and Android, and Voice is available in ChatGPT Work.

Sarvam agent workspace updates

Live voice input up to 5 minutes, one-command Slack connection, file sharing by link, artifact comments and faster batched agent outputs.

Kiro CLI 2.24.0 (and 2 earlier builds)

Session-wide tool approval and environment isolation (2.24.0); reasoning/effort controls in /model and cloud sessions (2.23.0).

Rollouts and Security Review

Two new bots: Rollouts tracks changes across environments; Security Review flags exploitable vulnerabilities in pull requests.

Local sandboxing in the Copilot app; configurable code reviews

Sandbox policies limit agent access to files, network and credentials; new code-review settings page and enterprise effort defaults.

Lovable: Claude models without API keys; new cloud connectors

App AI features can use Claude models without keys (GPT-6 Sol/Luna added 25 Sep); AWS, Azure, Google Cloud and Looker connectors.

Zed 1.21.0

SuperGrok sign-in, idle-sleep prevention while agents run, and BYOK for Claude Opus 5.5 and GPT-6 models.

Ground Google Docs prompts on Gemini Notebook sources

Docs can reference Gemini Notebook sources to ground AI drafting with inline citations.

Antigravity SDK supports local models

Offline agent workflows with local models such as Gemma 4 26B via LiteRT and OpenAI-compatible servers.

Runway in DaVinci Resolve; Runway MCP on Cursor and Grok

DaVinci Resolve plugin; Runway MCP on Cursor Marketplace and Grok Bot; Brand Kits and four Kling models via MCP.

Flashcards in ChatGPT

Interactive flashcard creation and practice on all plans.

Keyterm prompting on streaming speech-to-text

keyterms parameter now works on the streaming WebSocket STT endpoints with Saaras v4.

Also announced by Sarvam docs

Temporary rate-limit increase for Qwen 3.8 27B

Developer-tier limit for Qwen 3.8 27B raised from 450K to 750K tokens per minute for now.

Notion file uploads accept CAD/3D; higher MCP limits

File Upload API accepts .dwg, .dxf, .step, .stl and similar; MCP search/query limits raised to 20 calls per 10 s.

Jev as a judge

TypeSafe's Jev decision model can be used as a lower-cost evaluator judge.

Privacy Center; credit scores in Finances (US)

Central privacy-controls hub for consumer plans; US Plus/Pro users can link Experian credit report and VantageScore 3.0.

Effort mode, GPT-6 Astra and Skills marketplace

Perplexity Computer gains selectable effort levels, GPT-6 Astra, a skills marketplace, hybrid cloud/local compute on Mac and Side Chat.

Bolt training opt-out setting

New account setting to stop projects and chats being used for model training; default depends on region.

LLM Gateway adds Nemotron, DeepSeek V4.1 Flash, GLM 5.3 and 5.3 Flash

Four more models available through AssemblyAI's LLM Gateway.

n8n 2.40: agents callable from workflows and Slack

n8n 2.40 introduces agents built in n8n that can be called from workflows, Slack and other channels.

LM Studio 0.4.25

Adds the Splash engine on Apple Silicon M3 or newer with macOS 26.4+.

ChatGPT for Word; multiple accounts per plugin

ChatGPT inside Microsoft Word for drafting, summarising and revising on all plans; several Gmail/Calendar/Contacts accounts can be connected.

Tests for voice agents

Voice agents can be regression-tested with test cases, with API support for managing tests.

Enhance Frame Rate on Runway Dev

Convert videos to 24-120 fps; billed at 1 credit per 2 seconds.

Suno Studio MIDI improvements

More accurate MIDI transcription and new audio stems from a MIDI clip.

Deepgram India endpoint generally available

api.in.deepgram.com keeps data processing within India.

Kimi Web Search API; billing split change (day not given)

Web Search API added at /v1/tools/search; charges now deducted 50% cash and 50% vouchers. Page gives only the month.