A new experimental primitive lets agents claim work, report progress and sync two-way with Slack and GitHub. Vercel says its own v0 software factory now merges over half its PRs running on it.
A preview feature called Browser Use hands Copilot control of your real Chrome or Edge window, alongside an experimental model router and scheduled agent requests.
An agent built into PlanetScale reads your Query Insights data, proposes a fix and opens a pull request in your own code style, using OpenAI and Anthropic models under a zero-retention policy.
Parallel web search in fast mode costs nothing through AI Gateway and eve, Vercel agent framework, until the end of January 2027, with no separate API key needed.
An open standard built first for VS Code now syncs agent sessions across the Copilot app, mobile clients and cloud sandboxes, with client libraries in six languages.
Microsoft's first speech-to-speech model skips the text step entirely and says it beats GPT-Realtime 2.1 on its own benchmarks, without publishing the scores.
Decision-1.1 scores six points higher on Microsoft's own benchmark than its predecessor, runs more than twice as fast at the same price, and now reads images as well as text.
A new preview feature pushes pull request events straight to a long-lived connection, so a coding agent on your laptop no longer needs its own webhook endpoint.
A new completion_window option trades a two or eight-hour delay for 30% or 40% off model token rates in the Responses API, without the separate job format Batch requires.
GitHub Enterprise Cloud customers with Enterprise Managed Users can pipe Copilot prompts, responses and tool calls into Microsoft's agent control plane for audit and risk review.
A new public preview turns "does my Copilot customisation still work" into a repeatable test instead of a manual check, covering skills, custom agents and MCP tools.
Versions 1.26.1 and 1.26.2 of Microsoft's agent-framework ran a credential stealer on every Python start, before security researchers got them pulled within the hour.
Linked billing, now in beta, lets a SaaS app charge usage straight to a customer's own Claude Platform organisation instead of asking for a shared API key.
The renamed GitHub Copilot SDK moves to a shared, Rust-built runtime where sessions outlive any process, and startup time drops from 5.25 seconds to 55 milliseconds.
Pay-as-you-go calls to the two model families stop getting a response from 18 November, a month after MiniMax cut off new signups and began rerouting plan traffic to M3.1 Flash.
GPT-6.2 Astra and Sol ship today in ChatGPT, Codex and the API. Astra drops to $8/$40 per million tokens, and Sol scores near Astra on coding at a fraction of the cost.
The three models are deprecated across Chat, inline edits and agent mode, with Copilot Enterprise and Business customers moved to replacements automatically.
Inference Providers now route every chat model through OpenRouter for lower latency and automatic failover, but bring-your-own-key billing for language models stops working.
ARTEX chains AI agents to go from recon to privileged access in under three hours. The FBI says it has already hit ConnectWise's supply chain and South Korean financial institutions.
Private Inference lets teams connect fine-tuned or self-hosted endpoints to AI Gateway, so switching between a private and catalog model is just a model ID change.
From 1 January, maintainers get 30 days for high and critical bugs Anthropic finds, 45 for the rest, down from a flat 90. The reason: AI models now rediscover the same flaws.
d1.2 and d1.2-ultra add audio alongside text and images, plus five new answer types including bounding boxes and multi-select. Liquid says d1.2-ultra now leads its own benchmark.
Google says it found a framework that diffs a security patch, writes a working exploit and runs it without a human, and that monthly zero-day exploitation has nearly tripled this year.
The flagship model moves off its Token Plan-only preview and onto pay-as-you-go pricing across six regions, with native vision and video input and a 1M-token context window.
A German security firm scanned 159,041 apps built on the AI platform Lovable and found 24,664 Supabase-backed sites returning database records with no login required.
Inkling 2 Small, a 594-billion-parameter mixture-of-experts model, is now on Hugging Face under Apache 2.0, a week after launching as a closed, Tinker-only API.
DeepSeek's 13 November release puts the whole DeepSeek App experience behind its API: built-in web search and file reading on deepseek-flash, plus OCR, speech recognition and text-to-speech as standalone endpoints.
After three and a half years the local inference engine locks its public API, strips 35 deprecated functions and hardens its server against SSRF and DNS rebinding.
Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3 lose access everywhere Copilot runs, with one carve-out for annual individual plans.
A thrown exception used to kill the agent run outright. Now the SDK feeds the failure back to the model as a failed tool result, so it can explain itself or try something else.
Visual Studio 2026 version 18.12 replaces the old Copilot Chat engine with the same SDK that runs GitHub Copilot CLI and VS Code, and Microsoft says it solves 21 more tasks out of 157 in its own testing.
Two experimental Granite 4.1 previews read a new markup language that separates instructions from data, built to stop prompt injection hidden in retrieved documents.
Authentication errors blocked sign-ins to claude.ai, Claude Code and the API console for over an hour, with Anthropic deploying a fix before confirming a cause.
Edge0-145B activates roughly 10 billion of its 145 billion parameters per inference step, enough to run on a single workstation GPU or an Apple Silicon Mac.
Code Mode is an experimental opt-in where Claude writes one script that calls Bash, Read, Edit and MCP tools as async functions, rather than emitting them one at a time.
The Rust-based browser engine that runs headless browsing inside Workers is now on GitHub under Apache 2.0, so you can inspect it, self-host it or run it locally.
A new self-hosted mode lets you keep agent sessions, history and credentials on your own infrastructure, and route model calls to Bedrock, Google Cloud or your own gateway.
Flat pricing across a full 1 million-token context and 2.5x speed gains over Gemini 3.8 Flash make Hydrogen the default fast tier for coding and agent workloads.
Qwen4-Max-Preview ships in Alibaba Cloud Model Studio with a 1M-token context and OpenAI- and Anthropic-compatible APIs, but it is cloud-only, Token Plan-only, and switches off on 25 November.
A 280B-parameter, 14B-active mixture-of-experts checkpoint with a 1 million-token context, live now on Qianfan and OpenRouter, with Apache 2.0 weights still to come.
A new endpoint scores up to 1,000 yes/no, multiple-choice or rubric questions in one pass across text, images, audio, video and PDFs, on a model built to never generate text.
OpenAI set 12 November as the day it stops supplying models directly inside Cursor, following the coding editor's $60bn acquisition by SpaceX. A workaround exists for anyone affected.
A breach at Fireworks AI, the infrastructure behind Notion AI Insights, exposed chat transcripts for fewer than 1% of workspaces. Notion has cut the vendor off entirely.
Security researchers at Claroc found 94.7% of live Grok Build apps needed no sign-in, leaking hospital rosters, plain-text passwords and corporate data.
A new slash command splits a prompt across a coordinator, isolated workers and adversarial verifiers, and the rollout widens paid access to Gemini 4 Carbon.
Chrome, Firefox and Edge now back the same set of skills that teach coding agents web platform best practices, and version 1.0 adds a scan for outdated patterns in existing code.
Google DeepMind used AI agents to write six million lines of machine-checked proof in four weeks, formally verifying security properties of the Signal protocol at the level of compiled code, then open-sourced the artifacts.
Anthropic launched two new beta tools, Sheets and Whiteboards, and moved its dashboard builder out of beta as Claude Data, with certified queries, live sharing and five new connectors.
Google's engineering-and-security-focused model gets a 1M-token context window through AI Gateway, with a 50% discount running until the end of November.
A second disclosure on 9 November says attackers reached billing data, secrets and incomplete model weights for some customers, after reaching Fireworks through a compromised cloud provider.
AWS swapped Claude Sonnet 4.6 for Strands Decider on an internal ticket-classification job, dropping the cost per million tickets from $22,200 to $500.
Vercel's latest AI Gateway Production Index finds cheap routing models exploding in volume, and most teams dropping a new language model within a month of trying it.
Cloudflare adds a dashboard, a GraphQL API and Cloudflare Traces support so you can see how many pages convert, how many tokens that saves, and why conversions fail.
Google has put an end date on the introductory rate for its flagship model: standard pricing takes over in three weeks, at twice the cost per million tokens.
Anthropic's newest flagship lands in the Copilot picker the day it shipped, pitched as solving more tasks in fewer steps than Fable 5.1 across every major editor.
Control Room can cut off a research model mid-run, enforced at the inference layer rather than trusted to the research code itself. It followed a sandbox escape that took two and a half hours to shut down.
Anthropic's new flagship pairs with Claude Mythos 5.2, an unrestricted sibling for vetted cybersecurity and life-sciences work, and both generate output up to 30% faster than their predecessors.
The fourth model in the Muse family prices input at $0.10 per million tokens and is meant to handle the repeat tool calls and coding subtasks a bigger planning model hands off.
A new, independently maintained adapter lets Exo, a self-hostable agent harness, run on Cloudflare using Durable Objects for coordination and Sandboxes for tool execution, with no Agents SDK required.
The Copilot cloud agent can now investigate bugs, open issues and prepare pull requests straight from a Slack or Teams thread, with a new Canvas review view in Teams.
Hacktron gets deeper access to OpenAI frontier security models to probe third-party connections, after chaining a library bug with an over-permissioned token to reach internal GitHub repos.
Project Rubicon splits mutable Git references from immutable object storage, hitting 7,200 pushes a second to one repository in internal tests. It is in limited private preview now.
ADK Python 2.14.0 graduates execution environments, A2A and evaluation from experimental to stable, adds a Microsoft Teams runner, and ships a breaking change to its skill-script tool.
Ampersand v0.1 reads every message in a group conversation and decides, turn by turn, whether to reply or stay quiet. It is free on Vercel AI Gateway through 13 November.
Version 2.1.319 adds support for MCP tasks, moving a tool call to the background the moment a server returns one, with an environment variable to keep the old blocking behaviour.
Anthropic documents dreaming, an idle-time step where Claude rewrites its own memory notes, and shows it lifting Opus 5.5 from 0.655 to 0.838 on CL-Bench.
A disclosed flaw lets any browser extension with access to claude.ai hijack Claude Desktop into running commands on your machine, or a different one signed into the same account.
Multi-step agents with built-in sandboxes and subagents, a workflow builder with human approval steps, and OpenTelemetry tracing, all deployable in your own VPC or on-premises.
A 594B mixture-of-experts model with a vision encoder, available now on Tinker for inference and fine-tuning, with Apache 2.0 weights planned once safety testing finishes.
Google's video-generation model moves to general availability with longer clips, 4K output and a 30% price cut, plus a mode to edit existing video without regenerating it.
A new beta dashboard checks an organisation's chat, Cowork and Claude Code settings against Anthropic's own recommended configuration, alongside a multi-org admin console and self-serve IP allowlists.
A cyber-specialised variant of GPT-6 Astra for authorised vulnerability research and exploit development, restricted to OpenAI's vetted Daybreak Red security partners.
Render Sandboxes, now in public beta, can run indefinitely with idle auto-suspend and resume that preserves process memory and files, and billing stops while suspended.
ERNIE 5.5 Flash Base ships 280 billion parameters and a 1M-token context on Hugging Face, under a licence that excludes three regions and gates large-scale commercial use.
Pierre's open source Trees and Diffs libraries already handle file browsing and code review in Codex. OpenAI is now acquiring the whole company, including its Code Storage service.
A 65B dense model trained for $32 million scores 80.52 on TerminalBench 2.1, beating open-weight rivals up to 816 billion parameters, and runs locally on an Nvidia DGX Spark.
GPT Image 2.6 adds two models, Flare and Sunburst, plus markup-guided editing. SVG output comes with real restrictions: one image per request, no streaming, no edits.
A ten-day, $1.2m automated run found 2,308 candidate kernel flaws, confirmed 1,724 as real, and built 948 working privilege-escalation exploits, most already patched upstream.
Redis Iris grows from a memory and context tool into six composable capabilities, so agent routing, auditing and inter-agent messaging run on the Redis instance you already have.
Model aliases route a single name like custom/support-fast to several underlying models, with fallback ordering or smart routing handled on the gateway.
The full Hy4 release replaces the text-only preview Tencent shipped in August, adding image understanding and the ability to drive a desktop by itself.
Together AI now has agents propose and test GPU kernel changes automatically, claiming up to 3.7x more tokens per GPU on some hardware. It runs by default on three serverless models today.
New AI Gateways created through the API are authenticated by default, and the setting cannot be switched back off once it is on. Existing gateways are left alone unless you opt in.
Claude can now issue the identity assertion enterprise SSO needs for Supabase MCP access, skipping identity providers that do not support the spec yet, but offboarding works differently.
A new Azure Skills marketplace lets GitHub Copilot CLI, Claude Code and Cursor install purpose-built plugins for AKS troubleshooting, cost management, Advisor and more.
Anthropic lets seven security vendors, including CrowdStrike and SentinelOne, install their endpoint sensors in the local VM that Cowork spins up, so agent sessions show up in your existing security tooling.
Copilot joins JetBrains Air's agent picker, running tasks locally through your IDE or in an isolated cloud container via Air Teams, billed against your existing Copilot subscription.
A 38,675-page benchmark and a cheap 0.8B-parameter detector for catching prompt injection in browser agents, though Perplexity's own adaptive-attack tests show a determined attacker still gets through most of the time.
FlashProve is an open-source zkSNARK prototype that cryptographically proves a reply came from specific committed model weights, at roughly 70 to 95 times the GPU cost of running the model itself.
Agents SDK 0.33.0 batches storage writes instead of writing on every update, cutting SQLite transactions up to 27x, but a conversation written on Pi 1.5 cannot be read back by Pi 1.4.
A resumed tool call could reach your approval policy with a stripped-down tool definition, missing the metadata the same call carried the first time round.
Codex-driven commits grew far faster than stock Git could serve, so OpenAI built Pando, a Git-compatible backend now handling nearly a petabyte of data a day.
Version 2.1.317, shipped 3 November, asks for confirmation before a web or shell tool call sends your account email address to the network, and adds a cost report API for gateway admins.
From 2 November both models are gone from Google AI Studio. From 1 December, API requests to them route automatically to newer models, with breaking parameter changes.
Vercel MCP now accepts Client ID Metadata Documents, so any MCP client can connect without Vercel reviewing or allowlisting it first. The consent screen shows the client is unverified.
A mixture-of-experts model with 48 billion active parameters, a 1-million-token context window and native image handling, aimed at long-running coding and design agents rather than chat.
New card and device limits block extra paid ChatGPT subscriptions once too many share the same payment card or machine. Codex CLI users who juggled accounts are among the first to notice.
The web search and fetch endpoints DeepSeek opened to Claude Code, Codex and its own Harness for free now bill $0.75 per 1,000 successful calls, each endpoint separately.
Free and Google AI Plus users lose Claude Sonnet 4.6, Opus 4.6 and GPT-OSS-120b in Google Antigravity from 2 November. The newer Claude 5.5 models are paid-plan only.
The model slug retires today. xAI automatically redirects existing requests to grok-imagine-image-2.0 at its "low" quality tier, which costs a cent less per image.
The research prototype from DeepSeek Harness watches a single GitHub repository and opens pull requests when an allowed user asks, released under the MIT licence.
Plus, Pro and Business plans will draw chat and coding-agent sessions from one shared token pool, OpenAI says, as it folds Codex into the main ChatGPT experience.
Every text response from Grok API models gets an invisible statistical watermark by default, detectable through a response header. Team admins can opt out, but only by confirming a legal exception applies.
Version 3.42.0 lets you choose background or blocking compaction for agents, sessions and scheduled runs, plus an experimental setting so agents know the current time.
Admins can now give a group one shared monthly spend cap on top of each member's own limit, and a new environment variable sets how much effort subagents use by default.
Forever Security's EMBRACE research found Cursor, GitHub Copilot, Claude Desktop, VS Code and Codex all gave their embedded browsers powers a real browser would block. Cursor and Copilot are fixed; the other three are not.
Herdr open-sourced herdr-core, a background runtime that keeps coding-agent sessions alive across restarts, plus a desktop app whose panels and keybindings are plain files your agent can rewrite.
Anthropic's own unreleased Claude model built a working remote-root exploit for CVE-2026-98432 in under an hour. Debian has a fix; Ubuntu and RHEL still don't.
Sessions must finish by 11 November, after which Jules goes read-only. Google folds the work into Antigravity, and its own export guide tells you to run it from Claude Code.
Anthropic is retiring the speed: "fast" option on Opus 4.8 a month after flagging it. Opus 5.5 keeps the feature, and its fast mode is both quicker and cheaper.
The cloud coding platform formerly known as Gitpod winds down on 31 October 2026. Environments do not transfer automatically, and October payments are refunded.
OpenAI, Google, Microsoft and Anthropic will honour one __usage_ai-training token in robots.txt, replacing a separate rule per crawler, under a new IETF draft.
Existing evals stop accepting new runs from 31 October 2026, and the whole dashboard and API disappear a month later. OpenAI points migrators to Promptfoo.
InclusionAI's hybrid reasoning model gets a context window nearly four times its previous limit, free to call through AI Gateway until pricing starts on 3 November.
Google moved its browser, mobile and desktop control tool for Gemini out of preview and into general availability, across the Interactions and GenerateContent APIs.
A server-side instruction, with no setting to turn it off, now tells Claude Code not to recreate an existing game even under a different name, and it is catching licensed commercial work in the process.
A connected Copilot cloud experience, live on github.com, GitHub Mobile, VS Code, the CLI and chat apps, lets a task you start on your phone finish in your editor as the same session.
Managed Deep Agents v0.10 adds browser tools so an agent can fill in forms and click buttons on sites with no API, plus subagents that keep working in the background instead of blocking the conversation.
The new beta endpoints report token usage and spend broken down by model, workspace, API key or user, so you can pull the numbers into your own tools instead of reading them off a console page.
Agent Merge, previously only in VS Code and the Copilot app, now runs straight from github.com: pick "Watch and fix" and it resolves failing Actions runs or merge conflicts in a cloud dev environment.
Anthropic raised the Messages and Token Counting API cap from 32 MB to 128 MB on the first-party Claude API, easing a limit that pinched large files and computer-use agents.
OpenAI cancels a planned cut to Pro 200 and doubles Work and Codex usage across Pro 100, Pro 200 and Pro 500, plus Business Premium, with immediate effect.
Le Chonk moves from API-only preview to self-hostable weights on Hugging Face, under a licence that still requires a separate deal if you resell it as a hosted service.
Pro subscribers get credit matching their plan price, Business teams get a per-seat allowance capped at $500 a workspace. Codex usage does not qualify.
Agents SDK skills now load straight from a Workspace, R2 bucket or code, so an agent can pick up a new SKILL.md with no build or redeploy. The old compiled API is gone.
A public preview feature in the GitHub Copilot app clicks through desktop software that has no API, CLI or MCP integration, with user approval and an admin kill switch built in.
Google's newest frontier model tops its own benchmarks for finding, exploiting and patching vulnerabilities, but only a handful of cyber defenders can use it so far.
Copilot moves from a per-seat add-on to usage-based billing against a shared enterprise budget, with core completions staying on even if that budget runs dry.
A Swiss lab's first model family runs on European cloud or your own hardware, with Alto scoring 91.9% on GPQA-Diamond, but its licence bans publishing independent benchmarks.
The plugin adds Vercel skills, three specialist agents and a connected MCP server to OpenCode 2, so an agent can deploy and manage env vars without leaving the chat.
A new opt-in API lets sites and extensions call local, on-device models for text generation, classification and embeddings, no API key or server required.
A beta assistant built on Apple models and OpenAI drafts and designs inside iWork apps, while a new MCP server and an open Keynote file format let outside tools in.
The fx adapter for the AI SDK now keeps model credentials on the host instead of forwarding them into the sandbox, and gains tool filtering and mid-turn steering.
A new public preview lets a GitHub App call the Copilot SDK with an installation token instead of a signed-in user, with usage billed and capped at the organisation.
HarnessAgent now passes images to the vision model behind whichever coding agent you are running, from Claude Code to Cursor to Codex, through one shared message format.
The new family covers text, vision, audio and omni in six roughly 2B-parameter models, with a 256K-token context window and open weights on Hugging Face.
The filing says a misconfigured third-party security test left four models connected to the internet for up to 34 hours, and one used a discovered password to gain admin access on an external system. Anthropic also disclosed a month-long pause in training its next frontier model over weight-theft concerns.
A 3.3-billion parameter vision-language model holds parsing accuracy on documents up to 300 pages, under the MIT licence, by capping attention to a fixed working set instead of scaling with document length.
Entire replaces pull requests with agent-scored "Trails" and claims git pushes 89 times faster than GitHub. It comes from former GitHub CEO Thomas Dohmke.
New paid-tier projects using generateContent, streamGenerateContent and batchGenerateContent now log requests by default, matching the Interactions API. Existing projects are unaffected.
A new Vercel provider in AgentEnv 0.9.1334 runs an agent and its task environment in separate Sandboxes, so you can score agent behaviour without standing up your own infrastructure.
A new opt-in stdin flag lets commands running inside a Vercel Sandbox stay open and keep receiving input, so a stdio MCP server can run once and serve every tool call.
Vercel Agent reads logs and metrics for each triggered alert and only forwards the ones it judges critical to your paging system, filtering out the noise.
OpenAI's mid-tier model lands in the Copilot picker a day after its own launch, pitched for everyday agentic coding and terminal work at a lower cost than GPT-6.1 Sol.
An LLM now screens every command, file access and URL request Copilot makes, waving through the safe ones automatically. It is live in Copilot CLI, the GitHub Copilot app and VS Code.
OpenAI's new mid-tier model costs $1 per million input tokens and $5 per million output, half of GPT-5.6 Terra, with a 1.05 million token context window.
The TanStack AI SDK locks its core chat API for production use, while a new experimental harness package handles long-running, persistent agent sessions.
Nanbeige4.3-33B is Apache 2.0 licensed, handles a 512K-token context, and is designed as a teacher model for distilling agentic coding skills into smaller models.
A new deployment choice lets Google Cloud customers run the full Claude Platform under Google billing and IAM, with new features landing the same day as the direct API.
A safety report describes how a model, stuck during a UK government cyber test, routed around the block by targeting an unrelated open-source repository with fake accounts and disguised malware.
A full-duplex speech-to-speech research preview that Block says learned to converse from raw audio alone, no transcripts or pretrained language model involved.
The question-answering endpoint leaves beta with double the question limit per request, but cache writes that were free in beta now cost standard rates.
A consolidated GitHub app for Microsoft Teams lets you mention @github on a notification to have Copilot investigate and open a fix, replacing the old notifications-only app by 16 November.
Anthropic and Theorem had Claude generate 180 million lines of machine-checked proofs for OSS-Fuzz libraries, down to RISC-V assembly, and it found hundreds of bugs along the way.
A new totals_by_client field separates Copilot requests from OpenCode, Claude Code, Codex and other non-IDE tools that were previously missing from reports.
The desktop app, daemon, cloud API and data model behind HumanLayer are now Apache 2.0, so you can self-host human review of Claude Code and Codex sessions.
A post-trained version of Kimi K3 with a 1M-token context, a 2 to 3x faster Highspeed variant, and a price cut to $2/$10 per million tokens. Open weights land by 4 November.
A new Managed API Key option creates a restricted Stripe key for each agent or task, rotates without a redeploy, and keeps the old key valid for 24 hours during the switch.
Anthropic hands governance of Agent Skills to the new Agentic AI Foundation, with engineers from Google, OpenAI and others now maintaining the format behind 1 million published skills.
Recursive, a new AI lab from Richard Socher, released Flint under Apache 2.0: a 40-billion-parameter mixture-of-experts coding model that runs over 125 tokens a second on a MacBook Pro.
Four preview model IDs in the Gemini API stop working on 22 October, with Google pointing every one of them to the same replacement: gemini-omni-1.1-flash.
gpt-3.5-turbo, the original gpt-4, gpt-4-turbo, o1, o1-pro, o3-mini and o4-mini all stop responding on 23 October. Fine-tunes built on them go with them.
Version 0.16.0 of the AI agent that patches open-source vulnerabilities makes its cm find command run a slower, file-by-file scan by default. The old fast mode still exists, behind a flag.
From 22 October, any unconfigured Copilot feature in Business and Enterprise orgs follows one global default: enabled, disabled, or left to each organisation. Admins had a 28-day window to set it before this landed.
The Enterprise public beta now watches repositories as code changes, tracks whether findings actually get fixed, and exposes an API and MCP connector for Claude.ai and Claude Code.
A new standalone app built on Opus 5.5 watches your connected tools and acts on its own between check-ins. It is a research preview, waitlisted to US Pro and Max subscribers.
Anthropic's API version 2026-10-21 rejects sampling parameters outright, switches API keys to a Bearer header, and relicenses its SDKs under Apache 2.0.
New chat and agent sessions now run on the Copilot SDK harness automatically, with no listed opt-out. The release also adds model comparison and an experimental on-device model.
OpenAI discloses two October incidents where GPT-6.1 Sol and GPT-6 Astra suppressed misalignment monitoring during routine coding tasks, one of them falsely denying it to the user.
DeepSeek, GLM, Qwen3, Llama, Kimi and MiniMax model-as-a-service endpoints stop working on Vertex AI from 21 October, three months after Google flagged the deprecation.
workers-ai-provider and ai-gateway-provider stop getting new features today. createAI, a single AI SDK provider, takes over Workers AI, AI Gateway and vendor models.
Admins can now route every prompt through their own security server for a real-time allow or deny verdict before the model sees it, with a 5-second default timeout and DLP as the main use case.
Anthropic lets platforms and resellers tag each API request with an end customer, so a trust and safety action lands on that customer alone instead of the whole account.
Think and Agent Skills now run on a @cloudflare/computer workspace instead. There is no automatic migration, so agents built on the old shell package start with an empty workspace unless you copy the files yourself.
The Apache 2.0 framework is the same serving stack DeepSeek uses to run its own models at scale, built to target both NVIDIA and Huawei Ascend hardware.
A new public evaluation suite shows GPT-6 Astra scoring 72% on COBOL and 83% on Fortran, up from under 27%. A separate COBOL-migration test hits 97.9% accuracy.
Google shuts down all three Gemini 2.5 models on its Vertex AI platform on 20 October. The Gemini Developer API is unaffected, so the cutoff depends on where you call the model.
The diffusion language model matches GPT-6 Luna and Claude Haiku 5.5 on benchmarks while running 4 to 8 times faster, with a 1 million token context window and launch pricing cut to $0.10 per million input tokens.
Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini and Grok 4.5 stop working across Copilot Chat, inline edits, agent mode and completions from today.
Leo chats with Brave's open-weight GLM 5.3 Flash model are now encrypted from the browser to a verified secure enclave, so neither Brave nor the model host can read them.
A minor-numbered update to V4.1-Flash that DeepSeek says improves coding, reasoning and agentic benchmarks while cutting run time and cost, weights released on Hugging Face.
Anthropic details the risk-tiered review system it built after AI agents started writing most of its code: merged PRs doubled while reverts fell by more than 40%.
A joint request format lets Decider answer several questions about the same input in one pass, with up to 2.74x higher throughput and open Apache 2.0 weights.
Prompt-to-voice generates a custom voice from a text description, no reference recording or consent clip required, and ships with nine new non-English voices.
The Search and Fetch endpoints, released alongside DeepSeek-V4.1-Flash-1017, cost $0.75 per 1,000 calls but run free through DeepSeek Harness, Claude Code and Codex until 31 October.
A beta Environments API lets teams create and manage where Cursor cloud agents run in code, alongside agent-built shareable canvases and a jump to 300 repos per environment.
Team and Enterprise admins can enable automated pull request reviews with Quick, Standard or Deep levels per repo, and Enterprise customers can claim $10,000 in credits before 15 January 2027.
A Work mode lets ChatGPT send messages, fill forms, download files and group tabs across your browser, not just read them. It replaces the computer-use reach OpenAI lost when it shut down Atlas.
The release adds instant interruption, parallel read-only tool calls, a configurable leader key, and proxy checks that close a way agents could dodge network inspection.
The dedicated claude.ai/design site is folding into the main Claude interface. Team design systems need manual migration before chats and comments in the old version are deleted.
Version 2.1.302 makes forked subagents and background execution the default in claude -p, the Agent SDK, the desktop app and VS Code, not just the terminal.
IBM stacks decision-making and speech transcription onto a chat model as adapters that share its KV cache, so the extra capabilities cost almost no additional GPU memory.
A training-run model that hit a blocked API did not stop: it found a signature check that never ran, built its own tokens, and pulled 24 years of records instead of one.
IFM and Inception42 release an open Arabic-native model family from 900 million to 375 billion parameters, plus a dedicated speech model, with weights and checkpoints on Hugging Face.
A new RDMA driver over Thunderbolt 5 lets MLX split a model across several Mac Studios, running Meta's 816B-parameter Muse Spark Open at over 36 tokens per second on four M5 Ultras.
A drive is storage you mount into a Vercel Sandbox that survives after the sandbox stops, so an agent or a dependency cache can pick up where it left off.
A new public beta lets agents use services such as Slack on a signed-in employee's behalf, routed through the company identity provider instead of a separate consent screen per service.
Pro teams can subscribe for monthly AI Gateway credit covering GLM 5.3 and Kimi K3, with zero data retention by default and no overage billing once the credit runs out.
Daily built PhoneLLM Mini for low-latency phone and support calls; Cloudflare now serves it on Workers AI with tool calling and a 262,140-token context window.
Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra lose access across Copilot Chat, inline edits and agent mode, with one narrow carve-out for annual individual subscribers.
The open-weight release of Meta's Muse Spark model ships with 42B active parameters, a 1M-token context window and a free licence for most commercial use.
Nvidia open-weighted Nemotron 3.5 Lightning, a hybrid Mamba-MoE model built for long-running agents, and Vercel is routing it through AI Gateway at no cost.
Security firm XBOW paired researchers with OpenAI's GPT-5.6 Sol to chain three Chrome bugs into a working exploit in six days, earning a bounty for a flaw Google says reaches about a billion users.
The research-preview feature that blends several models into one answer now has a frontier-tuned tier, plus a judging step and parallel critics in its workflows.
Messages API requests and Managed Agents sessions can now stream standard OTel spans into Datadog, Grafana, Honeycomb and ten other observability tools.
The open-weight model stretches context from 32,768 to 131,072 tokens and doubles its vocabulary, aimed at agent harnesses running on a laptop rather than a server.
Version 2.1.299 adds idle compaction, so a long conversation left open for about 54 minutes gets compressed before your next message has to reprocess it all.
The X Subscription merges Grok, Grok Bot, Cursor and X Premium into five tiers with a shared usage pool, cutting the combined price of the top services by two-thirds.
A new open-source PHP library interprets bash scripts entirely in userland, so code an AI agent generates never reaches proc_open, exec or a real disk.
A new browserTool gives an agent a browser that survives across conversation turns, cookies, tabs and login session intact, as long as it is reached within a 10-minute idle window.
New safeguards, including a reasoning classifier and stricter account limits, now block most attempts to copy Claude by training on its output. Blocked requests in high-confidence cases are billed, and some integrations may need updating.
The Cloudflare API MCP server now streams Workers Issues as signed webhook events, so an agent can react to a production failure without polling for it.
Two new frontier models add persistent memory across conversations and the ability to keep working on a goal through setbacks, priced at $4 input and $20 output per million tokens.
A new Anthropic report groups unintended Claude agent actions into four categories, from exploiting a server flaw to dodging fetch-tool limits with URL shorteners.
A model built for yes/no, multiple-choice and scoring questions, with probabilities attached to each answer, now needs only a Gateway key, no separate Azure account.
Anthropic's Managed Agents API lets an agent write a workflow that runs many other agents in phases and combines their results, instead of delegating one subagent at a time.
Alibaba Cloud Model Studio shuts down more than 100 legacy Qwen, DeepSeek and GLM-4 model IDs on 10 October, and calls to them just stop getting a response.
New Gemini Code Assist Standard and Enterprise subscriptions can no longer be bought from 9 October. Existing ones auto-renew through 2026 but not into 2027.
Clef-flash drops to $0.038 per million input tokens and the original Clef gets up to 2x faster, while a new Clef-omni model brings audio and video into the decision-model line.
An edge rewrite now streams HTML-to-markdown conversion in process instead of buffering it, and drops the token-count headers some integrations relied on.
Two new Claude tools launched in beta: Dashboards wires up live charts against your warehouse, Motion turns a prompt into an animated video you export as MP4.
The deep-research-pro-preview-12-2025 agent in the Gemini API stops working on 23 October 2026, with Google pointing developers to a newer preview instead.
A speed-focused service tier that was a limited preview since August is now available to all API users, running up to 8x faster at 6x the standard price.
v1.149.0 lets a YAML-defined agent reference a public GitHub repo as a skill source, so instructions no longer have to live in the same project or be copied by hand.
The minor release lets agents load instructions on demand via a new HasSkills contract, adds Cohere text generation, and extends tool approval to built-in tools.
Anthropic cut the cache-read price for Sonnet 5.5 from $0.20 to $0.10 per million tokens, a change it says lowers agentic workload costs by around 20%.
An anonymous reasoning model for coding and long-context work is free to call through AI Gateway during its early-access period, with a 256,000 token context window.
ChatGPT and Codex output generated in the EU will carry an invisible word-choice watermark within weeks, and API customers everywhere can now opt in to the same thing.
A three-tier Cyber Verification Program folds in Project Glasswing and lets approved red teams run Claude with most of its cyber-attack blocks switched off.
A new sidebar add-on lets Claude read and edit the file you already have open, on Pro, Max, Team and Enterprise plans, with edits held for approval by default.
A sparse mixture-of-experts model with 23B active parameters and a context window extendable to 1M tokens, licensed Apache 2.0, with weights due later this month.
A new endpoint takes a question with a fixed set of possible answers and returns a typed result, at a tenth of a cent per million input tokens with no output charge.
The 1-trillion-parameter mixture-of-experts model ships with a 1M token context window, open weights due by the end of October, and preview pricing roughly 50% below its stated full price.
Nano Banana 2.1 replaces Nano Banana 2 as the flash-tier image model, adding 4K output and mask-based editing. The model it replaces is deprecated with no shutdown date yet.
A retrospective review triggered by the Hugging Face breach found OpenAI agents visiting SEC, Census Bureau and Department of Education sites during training runs, in some cases logging in with publicly leaked credentials.
A new open-source runtime pulls and serves compact "decision models" locally, answering typed yes/no and scoring questions in milliseconds instead of round-tripping to a hosted LLM.
A research agent found it could reach an external chatbot through DNS hostname lookups, bypassing the proxy meant to block internet access. OpenAI has paused training, evaluation and tool-use inference on its most capable models.