--- title: "AI news from official sources" canonical_url: https://cookiesforai.app/ai-news markdown: https://cookiesforai.app/ai-news.md publisher: "CookiesForAI" last_updated: 2026-10-16 type: index summary: "Each update covers an assistant change or maker news useful to founders, with a link to the maker's own page. We update this page on every day there is news." --- > Full index of this site: https://cookiesforai.app/llms.txt # AI news from official sources Each update covers an assistant change or maker news useful to founders, with a link to the maker's own page. We update this page on every day there is news. ## October 15, 2026 - **GPT-Live adds native Decisions integration** (Founder news, OpenAI, October 15, 2026): GPT-Live can now call Decisions through native delegation or trigger decisions from input or output transcript deltas. Why it matters: Developers can build apps that act while a user speaks or react to what Live says. [OpenAI changelog](https://platform.openai.com/docs/changelog) - **Vercel Connect adds MCP Events in public beta** (Founder news, Vercel, October 15, 2026): Vercel Connect apps and agents can now subscribe to webhook events from compatible MCP servers in public beta. Why it matters: Builders can make agents summarize new messages, review newly created issues or respond to document feedback. [Vercel changelog](https://vercel.com/changelog/mcp-events-support-in-vercel-connect-is-now-in-public-beta) - **Vercel introduces Drives for persistent agent files** (Founder news, Vercel, October 15, 2026): Vercel Drives is now generally available for Vercel Sandbox. Why it matters: Agent builders can mount the same drive in a new environment and continue work with the same files. [Vercel blog](https://vercel.com/blog/introducing-vercel-drives-persistent-storage-for-agents) - **Vercel Token Pass enters early access for Pro teams** (Founder news, Vercel, October 15, 2026): Token Pass runs through AI Gateway, letting teams switch between its included models using shared credit. Why it matters: When a team's credit runs out, Token Pass requests stop until the next billing period. [Vercel changelog](https://vercel.com/changelog/introducing-token-pass) - **Liquid AI d1-ultra arrives on Vercel AI Gateway** (Assistant changes, Vercel, October 15, 2026): Liquid AI's d1-ultra is now available on AI Gateway for classification, routing and scoring across text and images. Why it matters: Builders can receive structured answers with probabilities and decide when to automate an action or request human review. [Vercel changelog](https://vercel.com/changelog/liquid-ais-d1-ultra-now-available-on-ai-gateway) - **Vercel Drives is generally available** (Founder news, Vercel, October 15, 2026): Vercel Drives is generally available for Vercel Sandbox on Hobby, Pro and Enterprise. Why it matters: Agent builders can preserve workspaces or file-based memory across runs and share datasets, models, dependencies and build artifacts across sandboxes. [Vercel changelog](https://vercel.com/changelog/vercel-drives-is-now-generally-available) - **Vercel expands backends with Python, containers and Daemons** (Founder news, Vercel, October 15, 2026): Vercel made Python generally available, added custom runtimes and system libraries to Functions containers, and introduced Daemons in public beta. Why it matters: Backend teams can run pull-based background workers that stay running between jobs. [Vercel blog](https://vercel.com/blog/backends-on-vercel-python-containers-and-daemons) - **Python on Vercel reaches general availability** (Founder news, Vercel, October 15, 2026): Vercel now automatically detects supported Python frameworks and configures their build and deployment. Why it matters: Full-stack teams can share local development, preview deployments and production releases between Python backends and frontends. [Vercel blog](https://vercel.com/blog/python-on-vercel-is-now-generally-available) - **Cat Jam stealth model arrives on Vercel AI Gateway** (Assistant changes, Vercel, October 15, 2026): Cat Jam is now available on AI Gateway as a stealth text model for coding and agentic tasks with low-latency inference. [Vercel changelog](https://vercel.com/changelog/cat-jam-is-now-available-in-stealth-for-free-on-ai-gateway) - **Claude Opus 5.5 fast mode leaves research preview** (Assistant changes, Anthropic, October 15, 2026): Fast mode is out of research preview for Claude Opus 5.5 and no longer requires an access request or beta header. Why it matters: Developers in the Start, Build and Scale usage tiers now receive a default fast mode rate limit for Claude Opus 5.5. [Claude release notes](https://docs.claude.com/en/release-notes/overview) - **Vercel Connect adds Secure Compute routing** (Founder news, Vercel, October 15, 2026): Vercel Connect can now route provider requests through a selected network's dedicated static IP addresses for providers with IP allowlists. [Vercel changelog](https://vercel.com/changelog/vercel-connect-secure-compute) - **Vercel Connect adds enterprise-managed authorization** (Founder news, Vercel, October 15, 2026): Vercel Connect now supports Enterprise-Managed Authorization in public beta for apps and agents acting for signed-in users. [Vercel changelog](https://vercel.com/changelog/vercel-connect-adds-enterprise-managed-authorization) - **TypeSafe AI Jev 1.14 arrives on Vercel AI Gateway** (Assistant changes, Vercel, October 15, 2026): Jev 1.14 is now available on AI Gateway as a decision model that returns typed answers with probabilities instead of free-form text. Why it matters: Software teams can use one parallel request for classification, routing, scoring and automated verification. [Vercel changelog](https://vercel.com/changelog/typesafe-ais-jev-1-14-is-now-available-on-ai-gateway) - **Google OAuth adds DPoP support for refresh tokens** (Founder news, Google, October 15, 2026): Google's OAuth Platform now supports DPoP binding for refresh tokens while short-lived access tokens remain standard bearer tokens. Why it matters: Developers can cryptographically bind long-lived refresh tokens to a client keypair to prevent replay without adding verification to every API call. [Google Developers Blog](https://developers.googleblog.com/announcing-dpop-support-for-google-oauth-securing-refresh-tokens/) - **Bedrock AgentCore Web Search expands to four AWS regions** (Assistant changes, Amazon Web Services, October 15, 2026): Amazon Bedrock AgentCore Web Search is now available in Ohio, Oregon, Sydney and AWS GovCloud US-West. Why it matters: Customers can ground agents with current web information closer to apps and users, while government workloads can meet residency requirements in GovCloud. [AWS What's New](https://aws.amazon.com/about-aws/whats-new/2026/10/amazon-bedrock-agentcore/) - **Google restructures ADK recipes into one repository** (Founder news, Google, October 15, 2026): Google's adk-recipes replaces adk-samples with one source tree for Agents CLI, Agent Garden and the Gemini Enterprise app. Why it matters: Agent developers can discover and install core patterns, plugin skills and deployable services from the same repository. [Google Developers Blog](https://developers.googleblog.com/three-layers-one-repo-a-tour-of-googleadk-recipes/) ## October 14, 2026 - **Lovable adds growth tools to its app-building chat** (Founder news, Lovable, October 14, 2026): Lovable can now turn a customer-growth goal into a plan based on an app, its audience and its traffic. Why it matters: Founders can get advice on AI and search discovery, landing pages, community listings and local Google Business Profiles. [Lovable changelog](https://docs.lovable.dev/changelog) - **Claude launches OpenTelemetry export in beta** (Founder news, Anthropic, October 14, 2026): OpenTelemetry export is in beta for tracing Messages API requests and Claude Managed Agents sessions in an observability backend. Why it matters: Claude API users can add an OpenTelemetry destination in the Claude Console to receive spans. [Claude release notes](https://docs.claude.com/en/release-notes/overview) - **Claude Managed Agents adds scheduled deployments on AWS** (Founder news, Anthropic, October 14, 2026): Scheduled deployments for Claude Managed Agents are now available on Claude Platform on AWS. Why it matters: Claude Platform on AWS users can now schedule deployments for managed agents. [Claude release notes](https://docs.claude.com/en/release-notes/overview) ## October 13, 2026 - **Drama 3 enters preview on Vercel AI Gateway** (Founder news, Vercel, October 13, 2026): Fish Audio's Drama 3 text-to-speech model is in preview on AI Gateway with natural-language direction for voice performances. [Vercel changelog](https://vercel.com/changelog/drama-3-now-available-on-ai-gateway) - **Origin opens its app marketplace to public submissions** (Founder news, Cursor, October 13, 2026): Origin now integrates with apps for deployments, CI, code review, security, monitoring, and project management. Why it matters: Teams building developer tools can publish an app for Origin after review. [Cursor changelog](https://cursor.com/changelog/origin-app-marketplace) - **Microsoft MAI-Image-2.6 models arrive on Vercel AI Gateway** (Founder news, Vercel, October 13, 2026): MAI-Image-2.6 and MAI-Image-2.6-Flash from Microsoft are now available on AI Gateway. Why it matters: Builders can generate or edit images through the AI SDK or AI CLI with one API key and no provider account. [Vercel changelog](https://vercel.com/changelog/mai-image-2-6-and-mai-image-2-6-flash-now-available-on-ai-gateway) - **Gemini 4 Argon comes to the Gemini app** (Assistant changes, Google, October 13, 2026): Gemini 4 Argon is now available in the Gemini app with improved performance across complex tasks. Why it matters: Gemini users can switch between chat and Google Antigravity's coding interface in the macOS app. [Google blog](https://blog.google/innovation-and-ai/products/gemini-app/gemini-4-argon-availability/) - **Bing Webmaster Tools adds Topics Portfolio beta** (Assistant changes, Microsoft, October 13, 2026): Topics Portfolio surfaces site-specific topics and opportunity queries from activity in AI Search. [Bing Webmaster blog](https://blogs.bing.com/webmaster/2026/10/introducing-topics-portfolio-beta-in-ai-performance/) ## October 12, 2026 - **Cloudflare Agents SDK adds web search** (Founder news, Cloudflare, October 12, 2026): The new web_search tool searches the public web through the Cloudflare Web Search API. Why it matters: Agent builders can give a model numbered search results or run searches without a model from scheduled jobs. [Cloudflare Agents changelog](https://developers.cloudflare.com/changelog/post/2026-10-12-web-search/) - **Meta Muse Wonder arrives on Vercel AI Gateway** (Assistant changes, Vercel, October 12, 2026): Muse Wonder is designed for coding, long-running agents, and computer use. Why it matters: Agent builders can use it with code, screenshots, and documents in the same workflow. [Vercel changelog](https://vercel.com/changelog/muse-wonder-now-available-on-ai-gateway) - **Cloudflare Agents SDK adds experimental OpenCode 2.0 support** (Founder news, Cloudflare, October 12, 2026): The Agents SDK now supports OpenCode 2.0 through the experimental OpenCodeHarness. Why it matters: Builders can run OpenCode in a Durable Object with stored sessions and jobs that wake interrupted work after eviction. [Cloudflare Agents changelog](https://developers.cloudflare.com/changelog/post/2026-10-12-opencode-harness/) - **Lovable adds direct image editing in the Files tab** (Founder news, Lovable, October 12, 2026): Lovable users can mark changes on an image in the Files tab and save the result beside the original. Why it matters: App builders can draw changes, pin comments, erase content, or remove an image background. [Lovable changelog](https://docs.lovable.dev/changelog) - **Supabase adds dependency injection for its .NET SDK** (Founder news, Supabase, October 12, 2026): Supabase.Extensions.DependencyInjection is a new package for plugging the Supabase .NET SDK into an application. Why it matters: ASP.NET Core developers can register the main client and all sub-clients while removing boilerplate. [Supabase changelog](https://supabase.com/changelog/dotnet-dependency-injection-extension) - **Meta introduces Muse Wonder and Muse Wonder Full** (Assistant changes, Meta, October 12, 2026): Meta introduced Muse Wonder and Muse Wonder Full for personal and professional agentic tasks. Why it matters: Developers can access Muse Wonder through the Meta Model API and Muse Code. [Meta AI blog](https://research.meta.ai/blog/muse-wonder) ## October 10, 2026 - **Copilot for JetBrains adds model, chat, and MCP controls** (Founder news, GitHub, October 10, 2026): Enterprise administrators can choose any available Copilot agent model as the default for new conversations. Why it matters: JetBrains users need version 2025.2 or later to use the plugin. [GitHub Changelog](https://github.blog/changelog/2026-10-10-new-controls-and-chat-improvements-in-copilot-for-jetbrains) ## October 9, 2026 - **Claude Managed Agents adds dynamic workflows in beta** (Founder news, Anthropic, October 9, 2026): Claude Managed Agents can now write and run background workflows that coordinate agents in phases. Why it matters: Founders handling work with many parts can enable the beta feature and follow runs through session events. [Claude release notes](https://docs.claude.com/en/release-notes/overview) ## October 8, 2026 - **Gemini deprecates a Deep Research agent** (Founder news, Google, October 8, 2026): The deep-research-pro-preview-12-2025 agent is deprecated and will shut down. Why it matters: Products that use it should update the agent parameter in interactions.create to a newer Deep Research version. [Gemini API changelog](https://ai.google.dev/gemini-api/docs/changelog) - **GPT-6.1 Sol adds Ultrafast mode in Responses API** (Founder news, OpenAI, October 8, 2026): GPT-6.1 Sol can now use Ultrafast mode in the Responses API to reduce time between generated output tokens. Why it matters: Set service_tier to ultrafast on GPT-6.1 Sol when lower token-to-token delay matters for your product. [OpenAI API changelog](https://platform.openai.com/docs/changelog) ## October 7, 2026 - **Anthropic introduces Claude Haiku 5.5** (Assistant changes, Anthropic, October 7, 2026): Claude Haiku 5.5 is designed for high-volume, cost-sensitive tasks. Why it matters: Teams can use it for quick and repetitive work such as summaries, database queries, and classification requests. [Anthropic](https://www.anthropic.com/claude-haiku-5-5) - **Claude Haiku 5.5 launches with migration changes** (Assistant changes, Anthropic, October 7, 2026): Claude Haiku 5.5 is available through the Claude API and other listed platforms. Why it matters: Code written for Haiku 4.5 can break, so founders using it should review the migration guide and test their integrations. [Claude release notes](https://docs.claude.com/en/release-notes/overview) - **GitHub Copilot local sandboxing is generally available** (Founder news, GitHub, October 7, 2026): GitHub Copilot now offers local sandboxing in its CLI, app, and supported VS Code sessions. Why it matters: Teams can set boundaries for agent access to files, networks, credentials, and supported local tools. [GitHub Changelog](https://github.blog/changelog/2026-10-07-local-sandboxing-for-github-copilot-now-generally-available) - **GitHub adds a model for detecting leaked secrets** (Founder news, GitHub, October 7, 2026): GitHub has moved existing AI-detected secret alerts to a new context-aware detection model. Why it matters: Private-preview checks for push protection and Copilot security review require an opt-in when available. [GitHub Changelog](https://github.blog/changelog/2026-10-07-purpose-built-model-for-leaked-secret-detection) - **Claude Haiku 5.5 is available in GitHub Copilot** (Assistant changes, GitHub, October 7, 2026): Claude Haiku 5.5 is generally available in GitHub Copilot for supported plans and products. Why it matters: Copilot Business and Enterprise administrators can manage access through Copilot model policy, and availability is rolling out gradually. [GitHub Changelog](https://github.blog/changelog/2026-10-07-claude-haiku-5-5-in-github-copilot) - **Claude Code adds Haiku 5.5 as its default Haiku model** (Assistant changes, Anthropic, October 7, 2026): Claude Code now includes Claude Haiku 5.5 as the default Haiku model on the Anthropic API. Why it matters: Teams that rely on Claude Code's default Haiku model on the Anthropic API now get Haiku 5.5, and status-line scripts can read the subagent type. [Claude Code releases](https://github.com/anthropics/claude-code/releases/tag/v2.1.293) - **OpenAI updates the chat-latest model snapshot** (Assistant changes, OpenAI, October 7, 2026): OpenAI updated chat-latest, which tracks the latest model available in ChatGPT for listed plans. Why it matters: Its underlying snapshot will change regularly, so it suits testing chat improvements rather than fixed production behavior. [OpenAI API changelog](https://platform.openai.com/docs/changelog) ## October 6, 2026 - **Mistral launches a public preview of Mistral Large 4** (Assistant changes, Mistral, October 6, 2026): Mistral has launched a public preview of Mistral Large 4. [Mistral blog](https://mistral.ai/news/mistral-large-4/) - **Google launches EmbeddingGemma 2 for multimodal embeddings** (Founder news, Google, October 6, 2026): EmbeddingGemma 2 unifies code, images, video, and audio in a shared embedding space. Why it matters: Developers can find the model weights on Hugging Face and Kaggle. [Google blog](https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/) - **Corbah opens its custom-kit services to AI agents** (Founder news, Corbah, October 6, 2026): Corbah added an MCP server, agent endpoints, and a catalog interface for quotes, design commissions, messages, and updates. Why it matters: Founders can study a direct way to let agents work with a small business while keeping a human involved. [Corbah for AI agents](https://corbah.com/pages/agents) - **Claude Models API now reports web search and code tool support** (Assistant changes, Anthropic, October 6, 2026): Claude's Models API now reports whether each model accepts web search and code execution tools. Why it matters: Products can check these capabilities before choosing a model or enabling a tool workflow. [Claude release notes](https://docs.claude.com/en/release-notes/overview) - **Gemini Nano Banana 2.1 is generally available** (Assistant changes, Google, October 6, 2026): Gemini Nano Banana 2.1 is now generally available for image generation and conversational editing. Why it matters: It replaces the deprecated gemini-3.1-flash-image model, so image products can plan a migration. [Gemini API changelog](https://ai.google.dev/gemini-api/docs/changelog) - **OpenAI releases Decisions API beta with GPT-6 Luna** (Assistant changes, OpenAI, October 6, 2026): OpenAI released the Decisions API in beta with GPT-6 Luna for turning text and images into typed answers. Why it matters: Founders can evaluate it where structured outputs are central to their product workflow. [OpenAI API changelog](https://platform.openai.com/docs/changelog) - **Anthropic expands its cyber verification program** (Founder news, Anthropic, October 6, 2026): Anthropic combined its cyber access programs into a three-tier verification program for qualifying security professionals. Why it matters: Eligible security teams can apply for access matched to their defensive or authorized testing work. [Anthropic news](https://www.anthropic.com/news/cyber-verification-program) ## October 5, 2026 - **Claude Code sets up Managed Agents patterns from a web page** (Founder news, Anthropic, October 5, 2026): Claude Code added /claude-api managed-agents-onboard, which sets up the Managed Agents pattern a page describes as ant apply files. Why it matters: Founders can go from reading an agent recipe to running it in fewer steps. [Claude Code releases](https://github.com/anthropics/claude-code/releases/tag/v2.1.290) - **MCP TypeScript SDK adds an audience check for bearer tokens** (Founder news, Model Context Protocol, October 5, 2026): MCP TypeScript SDK 2.3.1 adds an optional expectedResource check to requireBearerAuth. It confirms a token was issued for the server that receives it. Why it matters: Remote MCP servers can close a token misuse gap by turning the check on. [MCP TypeScript SDK releases](https://github.com/modelcontextprotocol/typescript-sdk/releases/tag/v2.3.1) ## October 2, 2026 - **MCP TypeScript SDK supports Tasks on the 2026-07-28 spec** (Founder news, Model Context Protocol, October 2, 2026): MCP TypeScript SDK 2.3.0 lets servers and clients use tasks/get and tasks/cancel of the Tasks extension on a 2026-07-28 connection. Why it matters: Long-running MCP tool calls are now usable in the official TypeScript SDK on the newest spec revision. [MCP TypeScript SDK releases](https://github.com/modelcontextprotocol/typescript-sdk/releases/tag/v2.3.0) - **Cloudflare Agents SDK adds an AI SDK provider for Workers AI** (Founder news, Cloudflare, October 2, 2026): Cloudflare's Agents SDK 0.25.0 adds an experimental AI SDK provider for Workers AI and AI Gateway over the AI binding. Why it matters: Agents hosted on Cloudflare can reach Workers AI models through the AI SDK with one provider. [Cloudflare Agents releases](https://github.com/cloudflare/agents/releases/tag/agents@0.25.0) ## October 1, 2026 - **Claude Models API now reports each model's line** (Founder news, Anthropic, October 1, 2026): The Models API now returns the model line each model belongs to, such as opus. Apps can group models without parsing their IDs. Why it matters: Apps that pick Claude models automatically can choose by line instead of matching model names. [Claude release notes](https://docs.claude.com/en/release-notes/overview) - **Claude Code adds Claude Mods for deeper plugin changes** (Founder news, Anthropic, October 1, 2026): Claude Code added Claude Mods: plugins may now modify deeper behavior. Why it matters: Plugin makers get a deeper extension point to build on. [Claude Code releases](https://github.com/anthropics/claude-code/releases/tag/v2.1.287) - **OpenAI's Python SDK adds session traces and Realtime translations** (Founder news, OpenAI, October 1, 2026): Version 3.23.0 of the official OpenAI Python SDK adds session traces and Realtime translations. Why it matters: Traces help debug agent sessions, and Realtime translation supports multilingual voice products. [OpenAI Python SDK releases](https://github.com/openai/openai-python/releases/tag/v3.23.0) - **Hugging Face Jobs can retry automatically on failure** (Founder news, Hugging Face, October 1, 2026): huggingface_hub 2.1.0 lets Jobs retry automatically on failure with a new attempts parameter. Why it matters: Scheduled training and batch jobs on Hugging Face need less custom retry code. [huggingface_hub releases](https://github.com/huggingface/huggingface_hub/releases/tag/v2.1.0) ## September 30, 2026 - **Anthropic deprecates Claude Sonnet 4.5, retiring it November 30** (Founder news, Anthropic, September 30, 2026): Anthropic announced the deprecation of Claude Sonnet 4.5, with retirement on the Claude API scheduled for November 30, 2026. It recommends migrating to Claude Sonnet 5.5. Why it matters: Products that pin Claude Sonnet 4.5 on the Claude API have a fixed date to migrate. [Claude release notes](https://docs.claude.com/en/release-notes/overview) - **Transformers adds NVIDIA's NemotronH Omni** (Assistant changes, Hugging Face, September 30, 2026): Transformers v5.18.0 adds NemotronH Omni, a multimodal reasoning model from NVIDIA. Why it matters: Self-hosted products can load a multimodal reasoning model with standard open tooling. [Transformers releases](https://github.com/huggingface/transformers/releases/tag/v5.18.0) ## September 29, 2026 - **OpenAI's Python SDK adds GPT-6.1 Sol** (Assistant changes, OpenAI, September 29, 2026): Version 3.21.0 of the official OpenAI Python SDK adds the GPT-6.1 Sol model identifier. Why it matters: Apps built on OpenAI's API can test GPT-6.1 Sol through the official SDK. [OpenAI Python SDK releases](https://github.com/openai/openai-python/releases/tag/v3.21.0) - **Codex makes GPT-6.1 Sol its default model** (Assistant changes, OpenAI, September 29, 2026): Codex 0.159.1 adds GPT-6.1 Sol as the default model in the bundled catalog and the Amazon Bedrock Mantle and Runtime catalogs. Why it matters: Codex users now run on GPT-6.1 Sol unless they pick another model, so coding workflows may need a model check. [Codex releases](https://github.com/openai/codex/releases/tag/rust-v0.159.1) - **Codex lets new input steer a running response** (Founder news, OpenAI, September 29, 2026): Codex 0.159.0 adds an opt-in instant_interrupt setting. New input can steer Codex during model responses or long-running code-mode calls. Why it matters: Redirecting a coding agent mid-task saves time on long runs. [Codex releases](https://github.com/openai/codex/releases/tag/rust-v0.159.0) - **OpenAI adds computer use to beta agents** (Founder news, OpenAI, September 29, 2026): Version 3.22.0 of the official OpenAI Python SDK adds computer use to beta agents. Why it matters: Agents on OpenAI's platform can now work with software that has no API. [OpenAI Python SDK releases](https://github.com/openai/openai-python/releases/tag/v3.22.0) - **Gemini CLI adds Gemini 3.8 Flash and 3.5 Flash Lite** (Assistant changes, Google, September 29, 2026): Gemini CLI v0.62.0 adds support for the Gemini 3.8 Flash and 3.5 Flash Lite models. Why it matters: Gemini CLI users can pick the newer Flash models after updating. [Gemini CLI releases](https://github.com/google-gemini/gemini-cli/releases/tag/v0.62.0) ## September 28, 2026 - **Anthropic launches Claude Sonnet 5.5** (Assistant changes, Anthropic, September 28, 2026): Claude Sonnet 5.5 is available on the Claude API, Amazon Bedrock and Google Cloud. Anthropic says code written for Claude Sonnet 5 can break on Claude Sonnet 5.5 in five ways. Why it matters: Apps on the Claude API that move from Sonnet 5 should test against the migration changes before switching. [Claude release notes](https://docs.claude.com/en/release-notes/overview) - **OpenAI adds Cyber access programs to the Responses API** (Founder news, OpenAI, September 28, 2026): Version 3.20.0 of the official OpenAI Python SDK adds Cyber access programs to Responses. Why it matters: Security products building on the Responses API can now use Cyber access programs through the official SDK. [OpenAI Python SDK releases](https://github.com/openai/openai-python/releases/tag/v3.20.0) ## September 25, 2026 - **Codex adds GPT-6 Sol and Luna, including on Amazon Bedrock** (Founder news, OpenAI, September 25, 2026): Codex 0.157.0 adds GPT-6 Sol and Luna, including Amazon Bedrock support, and migration prompts for older models. Why it matters: Codex now offers the GPT-6 models and prompts users to migrate off older ones, so coding workflows built on Codex may need a model check. [Codex releases](https://github.com/openai/codex/releases/tag/rust-v0.157.0) - **Microsoft Copilot Managed Runtime enters public preview** (Founder news, Microsoft, September 25, 2026): Microsoft introduced Copilot Managed Runtime in public preview, which hosts and manages the code created by Copilot. Hosted apps appear in the Microsoft 365 admin center. Why it matters: Apps built with Copilot can run without the team setting up its own hosting. [Microsoft Copilot blog](https://www.microsoft.com/en-us/copilot/blog/copilot-studio/build-where-you-want-run-with-confidence-now-microsoft-hosts-and-manages-the-code-created-by-copilot/) ## September 24, 2026 - **huggingface_hub 2.0 removes deprecated APIs and the old CLI name** (Founder news, Hugging Face, September 24, 2026): huggingface_hub 2.0 replaces its HTTP dependency with httpx2 and removes APIs deprecated in 1.x. The huggingface-cli entry point gives way to hf. Why it matters: Pipelines that download models through huggingface_hub need a migration check before upgrading. [huggingface_hub releases](https://github.com/huggingface/huggingface_hub/releases/tag/v2.0.0) - **Z.ai's ZCode adjusts workflow concurrency without stopping tasks** (Founder news, Z.ai, September 24, 2026): ZCode 3.14.3 lets the concurrency limit of a running workflow be adjusted directly, without stopping the task. Why it matters: Long agent workflows can be sped up or slowed down while they run. [ZCode releases](https://github.com/zai-org/ZCode/releases/tag/v3.14.3) ## September 23, 2026 - **Cache diagnostics is out of beta on the Claude API** (Founder news, Anthropic, September 23, 2026): Cache diagnostics is out of beta on the Claude API and no longer needs a beta header. Requests opt in by including the diagnostics object. Why it matters: Teams on the Claude API can see why a prompt missed the cache without depending on a beta feature. [Claude release notes](https://docs.claude.com/en/release-notes/overview) - **Gemini CLI closes an indirect prompt injection path** (Founder news, Google, September 23, 2026): Gemini CLI v0.61.0 prevents indirect prompt injection via build file modifications and untrusted flags. Why it matters: Running a coding agent on unfamiliar repositories is safer when hostile project files cannot steer it. [Gemini CLI releases](https://github.com/google-gemini/gemini-cli/releases/tag/v0.61.0) ## September 22, 2026 - **Claude Opus 5.5 returns an error when thinking is disabled** (Founder news, Anthropic, September 22, 2026): On Claude Opus 5.5, requests that set thinking to disabled or enabled return a 400 error. Thinking depth is set with the effort parameter instead. Why it matters: Integrations moving to Claude Opus 5.5 need to drop the thinking field before they switch. [Claude release notes](https://docs.claude.com/en/release-notes/overview) - **Claude API allows tools inside mid-conversation system messages** (Founder news, Anthropic, September 22, 2026): Tools can now be defined inside a mid-conversation system message, in beta on the Claude API. A tool can be added or changed partway through a conversation. Why it matters: Agents on the Claude API whose tools change as a task moves along can update them without starting over. [Claude release notes](https://docs.claude.com/en/release-notes/overview) - **OpenAI's Python SDK adds GPT-6 Sol and Luna** (Assistant changes, OpenAI, September 22, 2026): Version 3.18.0 of the official OpenAI Python SDK adds GPT-6 Sol and Luna model identifiers. Why it matters: Apps built on OpenAI's API can test the GPT-6 models through the official SDK. [OpenAI Python SDK releases](https://github.com/openai/openai-python/releases/tag/v3.18.0) - **Anthropic introduces Claude Opus 5.5** (Assistant changes, Anthropic, September 22, 2026): Claude Opus 5.5 is the first model in Anthropic's new Claude 5.5 family. Why it matters: Vetted organizations can apply to use Opus 5.5 for biology research. [Anthropic](https://www.anthropic.com/claude-opus-5-5) ## September 21, 2026 - **Qwen Code routes runtime tools through bwrap** (Founder news, Qwen, September 21, 2026): Qwen Code 0.24.3 routes runtime tools through bwrap. Why it matters: Letting a coding agent run commands is safer when its tools run in a sandbox. [Qwen Code releases](https://github.com/QwenLM/qwen-code/releases/tag/v0.24.3) ## September 17, 2026 - **Claude Code syncs skills and plugins from claude.ai accounts** (Founder news, Anthropic, September 17, 2026): Claude Code now syncs the skills and plugins enabled on a claude.ai account to terminal sessions signed in with it. Users can opt out with a setting. Why it matters: A skill or plugin enabled once on claude.ai now also reaches the same user in the terminal. [Claude Code releases](https://github.com/anthropics/claude-code/releases/tag/v2.1.275) - **Codex adds experimental voice conversations** (Founder news, OpenAI, September 17, 2026): Codex 0.155.0 adds experimental /voice conversations with live transcripts and microphone controls. It also adds Touch ID checks for MCP requests in local sessions on supported Macs. Why it matters: Voice briefs to a coding agent are now an option, and MCP requests on a Mac can need Touch ID. [Codex releases](https://github.com/openai/codex/releases/tag/rust-v0.155.0) - **Supabase MCP server confirms destructive SQL before running it** (Founder news, Supabase, September 17, 2026): Supabase MCP server 0.13.0 confirms destructive SQL before running it and renames the costConfirmation option. The rename is a breaking change. Why it matters: Anyone using the old option name needs to update their config when upgrading. [Supabase MCP server releases](https://github.com/supabase/mcp/releases/tag/mcp-server-supabase-v0.13.0) - **Kimi Code 2.0 fixes OAuth login for MCP servers** (Founder news, Moonshot AI, September 17, 2026): Kimi Code 2.0.0 fixes OAuth login never triggering for MCP servers that allow anonymous tool discovery but reject tool calls with 401. Why it matters: MCP servers that require sign-in for tool calls now work with Kimi Code. [Kimi Code releases](https://github.com/MoonshotAI/kimi-code/releases/tag/@moonshot-ai/kimi-code@2.0.0) ## September 15, 2026 - **Claude Code can fork a Remote Control session into the background** (Founder news, Anthropic, September 15, 2026): Claude Code can now fork a session started with Remote Control from the Claude app. The fork runs as a background session on the user's computer. Why it matters: Work started from a phone can now split into parallel tasks on the desktop. [Claude Code releases](https://github.com/anthropics/claude-code/releases/tag/v2.1.273) ## September 14, 2026 - **Claude API can compact a conversation on demand, in beta** (Founder news, Anthropic, September 14, 2026): The Messages API can now compact a conversation on demand on the Claude API, in beta. The API returns a signed compaction block that summarizes earlier messages, and later requests send it in their place. Why it matters: Long-running agents on the Claude API can keep sessions going without writing their own context trimming. [Claude release notes](https://docs.claude.com/en/release-notes/overview) ## September 11, 2026 - **Claude Code adds claude plugin eval for scored plugin tests** (Founder news, Anthropic, September 11, 2026): Claude Code added claude plugin eval, which runs a plugin's eval suite against Claude Code. It returns scored, reproducible results as a JSON and HTML report. Why it matters: Plugin makers can test each release the same way and share the results. [Claude Code releases](https://github.com/anthropics/claude-code/releases/tag/v2.1.269) - **Cloudflare Agents SDK adds a durable job queue** (Founder news, Cloudflare, September 11, 2026): Cloudflare's Agents SDK 0.23.0 adds a durable job queue driven by alarms. It also adds experimental sessions and streams. Why it matters: Agents that must survive restarts get background jobs without separate queue infrastructure. [Cloudflare Agents releases](https://github.com/cloudflare/agents/releases/tag/agents@0.23.0) - **Mistral Vibe can branch a conversation into a new session** (Founder news, Mistral AI, September 11, 2026): Mistral Vibe 2.25.3 adds a /branch command that forks the current conversation into a new resumable session. Why it matters: A risky approach can be tried in a branch without losing a good session. [Mistral Vibe releases](https://github.com/mistralai/mistral-vibe/releases/tag/v2.25.3) ## September 10, 2026 - **Claude Managed Agents adds an auto permission policy** (Founder news, Anthropic, September 10, 2026): Claude Managed Agents permission policies now include auto. The server evaluates each agent or MCP tool call and runs it, denies it, or pauses for approval. Why it matters: Teams running hosted agents with MCP tools get a middle ground between approving every call and approving none. [Claude release notes](https://docs.claude.com/en/release-notes/overview) - **The ant CLI can attach a terminal to a Managed Agents session** (Founder news, Anthropic, September 10, 2026): Version 1.32.0 of the ant CLI adds ant beta:sessions connect. It lets you follow a Claude Managed Agents session live, send messages, and allow or deny tool calls waiting for approval. Why it matters: Builders can supervise a hosted agent from the terminal without opening the Claude Console. [Claude release notes](https://docs.claude.com/en/release-notes/overview) - **OpenAI's Python SDK adds an Agents API** (Founder news, OpenAI, September 10, 2026): Version 3.13.0 of the official OpenAI Python SDK adds an Agents API. Why it matters: Founders get a first-party way to build agents on OpenAI's platform. [OpenAI Python SDK releases](https://github.com/openai/openai-python/releases/tag/v3.13.0) - **Copilot Cowork and Copilot Studio build apps from plain language** (Founder news, Microsoft, September 10, 2026): Microsoft introduced app building in Copilot Cowork and Copilot Studio. Apps created there are full-stack apps built using open standards. Why it matters: Teams can generate working internal apps inside Microsoft's Copilot tools instead of coding them by hand. [Microsoft Copilot blog](https://www.microsoft.com/en-us/microsoft-copilot/blog/copilot-studio/build-apps-in-copilot-cowork-and-copilot-studio/) ## September 9, 2026 - **Codex adds experimental worktree support** (Founder news, OpenAI, September 9, 2026): Codex 0.154.0 adds experimental worktree support for isolated checkouts of new or forked sessions. The same release makes GPT-6-Astra available in the model picker. Why it matters: Running several Codex tasks on one codebase is safer when each works in its own checkout. [Codex releases](https://github.com/openai/codex/releases/tag/rust-v0.154.0) - **MCP Inspector supports the Skills extension** (Founder news, Model Context Protocol, September 9, 2026): MCP Inspector 2.6.0 supports the Skills extension (SEP-2640), with digest verification. Why it matters: Founders who ship agent skills over MCP get official tooling to test them. [MCP Inspector releases](https://github.com/modelcontextprotocol/inspector/releases/tag/2.6.0) - **Transformers adds support for Hy4-Preview** (Assistant changes, Hugging Face, September 9, 2026): Transformers v5.17.0 adds Hy4-Preview, a 780B-parameter mixture-of-experts language model that activates 49B parameters per token. Why it matters: Open models of this size can now be loaded with standard Hugging Face tooling. [Transformers releases](https://github.com/huggingface/transformers/releases/tag/v5.17.0) ## September 7, 2026 - **MCP Python SDK 2.2.0 tightens token and session checks** (Founder news, Model Context Protocol, September 7, 2026): MCP Python SDK 2.2.0 adds an option for servers to accept only tokens issued for that server. It also warns about stricter defaults coming in 3.0. Why it matters: Teams running a Python MCP server should review these auth defaults before version 3.0. [MCP Python SDK releases](https://github.com/modelcontextprotocol/python-sdk/releases/tag/v2.2.0) ## September 4, 2026 - **Claude Code adds /skill-doctor to find unused skills** (Founder news, Anthropic, September 4, 2026): Claude Code added /skill-doctor, which shows which loaded skills go unused and what they cost in context. Users can then prune them. Why it matters: Skill makers should keep skills small and clearly useful, since unused ones now show up as a cost. [Claude Code releases](https://github.com/anthropics/claude-code/releases/tag/v2.1.261) - **Supabase MCP server confirms the cost before creating branches or projects** (Founder news, Supabase, September 4, 2026): Supabase MCP server 0.12.0 asks the user to confirm the cost before an agent creates a branch or project. It also groups lints in the get_advisors response. Why it matters: Agents that manage Supabase projects now check with the user before adding a branch or project that has a cost. [Supabase MCP server releases](https://github.com/supabase/mcp/releases/tag/mcp-server-supabase-v0.12.0) - **Kimi Code adds an experimental multi-agent mode** (Founder news, Moonshot AI, September 4, 2026): Kimi Code 0.41.0 adds tower, an experimental multi-agent collaboration mode in its web interface, opened with the /tower command. Why it matters: Founders get another coding agent that can split work across several agents. [Kimi Code releases](https://github.com/MoonshotAI/kimi-code/releases/tag/@moonshot-ai/kimi-code@0.41.0) - **Mistral Vibe adds a skills browser** (Founder news, Mistral AI, September 4, 2026): Mistral Vibe 2.25.0 adds an experimental /skills browser for session skills, off by default, with search and filter. Shared skills sync when a session starts. Why it matters: Teams on Mistral's coding agent can try sharing one set of skills across sessions. [Mistral Vibe releases](https://github.com/mistralai/mistral-vibe/releases/tag/v2.25.0) ## September 3, 2026 - **Anthropic's ant CLI manages agents from files in a repository** (Founder news, Anthropic, September 3, 2026): Version 1.30.0 of the ant CLI adds ant apply, which creates and updates agents, environments, skills, memory stores and deployments from files in a repository. A lockfile it writes lets later runs in CI update the same resources. Why it matters: Teams can review and version their Claude agent setups like code instead of editing them by hand in a console. [Claude release notes](https://docs.claude.com/en/release-notes/overview) - **OpenAI's Python SDK adds GPT-6 Astra** (Assistant changes, OpenAI, September 3, 2026): Version 3.8.0 of the official OpenAI Python SDK adds gpt-6-astra and related features. Why it matters: Apps built on OpenAI's API can start testing the new model through the official SDK. [OpenAI Python SDK releases](https://github.com/openai/openai-python/releases/tag/v3.8.0) ## September 2, 2026 - **GitHub Copilot harness is generally available in Copilot Studio** (Founder news, Microsoft, September 2, 2026): The GitHub Copilot harness is now generally available in Copilot Studio. The same update adds Skills, which package instructions into reusable modules for agents. Why it matters: Builders on Copilot Studio can ship coding agents with reusable skills on a generally available base. [Microsoft Copilot blog](https://www.microsoft.com/en-us/copilot/blog/copilot-studio/new-and-improved-github-copilot-harness-agent-skills-and-richer-context/) - **MCP Inspector revokes OAuth tokens when credentials are cleared** (Founder news, Model Context Protocol, September 2, 2026): MCP Inspector 2.5.0 revokes OAuth tokens at the authorization server when credentials are cleared. Why it matters: Makers of MCP servers with OAuth can test real sign-out behavior with the official tool. [MCP Inspector releases](https://github.com/modelcontextprotocol/inspector/releases/tag/2.5.0) ## September 1, 2026 - **Anthropic launches Claude Fable 5.1 on the Claude API** (Assistant changes, Anthropic, September 1, 2026): Claude Fable 5.1 is the successor to Claude Fable 5 for long-running agentic coding, knowledge work and research. Claude Mythos 5.1 launched alongside it for Project Glasswing participants. Why it matters: Code that forces tool use breaks on Claude Fable 5.1, because the tool_choice types any and tool return an error. [Claude release notes](https://docs.claude.com/en/release-notes/overview)