What changed in AI tools

Idle 200K+ token chats are now compacted before the one-hour prompt cache expires, and a new notice says standalone Claude Design (claude.ai/design) closes on 14 Dec 2026.

The editing model keeps products and characters consistent, takes up to 6 references, and is live on krea.ai and partner APIs.

Suggested replacements are Opus 5.5 and GPT-6.1 Sol or GPT-6 Luna; individual annual-plan subscribers keep Opus 4.8.

Haiku 5.5 is in Power Mode's model selector, and enterprises can run the full platform or published apps in their own cloud.

Probabilistic forecasting in four sizes (5M–209M parameters), CPU-efficient and ready for zero-shot use.

Moonshot's open-weight model has a 1M-token context and experimental support for Pro and higher in Kiro's US geography.

Create issues, open PRs and investigate bugs across repos without switching conversations.

Copilot explores the codebase, proposes a fix, reruns CodeQL to confirm it closes the alert, then opens a draft PR; requires Code Security or Advanced Security.

The feature generates music and sound effects from a video automatically.

Cloud agent environments now support up to 300 repositories, and large ones load 5× faster.

Some Grok requests failed with capacity errors from 3:25 pm on 15 Oct to 10:20 pm PDT, when Cursor declared them working normally.

The release also adds a leader key in /keymap and a /daemon restart, and turns on Daybreak by default for eligible accounts.

OpenAI's status page flagged widespread feature degradation, with Login among the affected components, before all services recovered.

Over a month of updates, runners learned to serve many directories, create Git worktrees, be shared with a workspace, start from the macOS app and expose their desktop.

Lovable builds a growth plan and can run campaigns on Google, YouTube, Facebook and Instagram from the build chat.

Reviews now finish about 3× faster (6–10 minutes) with per-repo Quick/Standard/Deep levels; eligible Enterprise orgs can claim at least $10,000 in credits by 15 Jan 2027.

Also better custom Bedrock models and lower memory and CPU use in long agent threads.

October builds bring first-class support for Kiro CLI and Muse Code, Codex cache-write token reporting and a Markdown raw/rendered toggle.

Both deployment targets are in private preview.

OS-enforced limits stop agent-run commands from reaching local files and credentials, in the IDE and CLI.

One asset can be written for a whole buying committee.

The multi-model orchestrator now draws on Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, sets effort per step, and stays in research preview.

The release also improves /ask answers (expand, copy, reopen) and fixes command hooks and MCP status sync.

xAI says the speech-to-speech model ranks #1 on Artificial Analysis's Speech-to-Speech Index (86.9) with under-one-second latency.

It is billed at provider list pricing under usage-based billing, with a gradual rollout across VS Code, the CLI, the coding agent and other IDEs.

The OpenCode-based CLI can read Jira issues, double-counted Gemini usage and cost data is corrected, and managed models get the latest GPT models.

As of 14 Oct users must pick another model; Sonnet 5.5 is the current Sonnet.

A remote MCP server puts the Kilo platform API in front of any MCP-compatible client.

Instant voice clones need no training and are free during preview, and heygen-voice-1-turbo speaks them faster and cheaper; public voices and previews followed.

External agents connect over MCP with admin controls, Notion agents reach 150+ tools, a Skills API syncs skills to other agents, and version history is upgraded.

A widespread issue caused errors across Cursor from 5:22 pm to 8:07 pm PDT on 12 Oct; Cloud Agents were the last to recover.

Admins can lock extensions to a private marketplace, and the model picker remembers your last variant per model family.

Vercel reported more build and deployment failures, then delayed deployment events, until the backlog cleared at 1:11 pm PDT on 12 Oct.

On Free, Pro and Business, usage details now split credits between daily build credits, the daily chat allowance, monthly grants and general credits.

EU inference covers building and published apps; some models aren't available and workspaces use more credits while it is on.

Elevated 500s began at 4:51 am PDT on 12 Oct and core features recovered by 5:41 am, but notification delivery wasn't fully mitigated until 1:16 pm PDT.

Apps from Cloudflare, Linear, Sentry, Socket and others cover deploys, CI, review, security and monitoring, and the marketplace is open to public submissions.

Puck MCP exposes two tools, puck and orbs, so you can hand work to Amp and check running orbs from other MCP clients.

New first-class GitHub webhooks let an agent subscribe to a PR's comments, reviews and CI, or spawn a new orb for every new issue or PR.

V3 brings the IDE's agent harness to the terminal and adds cloud sessions, spec mode and tangent mode.

From 7:30 am to 1:06 pm PDT on 9 Oct multiple services were unavailable; Augment blamed permissions after a recent infrastructure migration.

Extensions can supply AI models, Send Element/Window to AI share screen context, and File Search is up to 6× faster.

The Rust CLI gains first-run onboarding with keyring storage, /teleport to Vibe Code Web, /voice and /proxy-setup.

Haiku 5.5 reached Pro and above, local sandboxing went GA at no extra cost, and the Copilot app can use separate accounts for license and repos.

Add brand assets once and every new creation pulls them in.

2.1.293 makes Claude Haiku 5.5 the default Haiku ($0.10/$0.50 per Mtok, 1M context); hooks gain onFailure: "block" and the Agent tool gains an effort parameter.

Agent builds run in isolated local sandboxes (Microsoft MXC, NVIDIA OpenShell); chats can now read and task your other projects.

On the alpha site, V8.2 jobs can be rerun with thinking to fix missed prompt details, and folders can be shared with collaborators or viewers.

Publishing to a whole workspace or to groups still requires Business or Enterprise.

Still 2.5B active parameters under Apache 2.0, but reinforcement learning means it can now explore a repo, edit files and check its own changes.

Claude Dashboards and Claude Motion enter beta at the same time.

The app combines agentic coding and data science workflows with built-in management of local models and conda environments.

Anthropic calls Haiku 5.5 its cheapest, fastest and most capable small model; Max and Team plans now include monthly API credits for the Claude Platform.

OpenAI's status page listed most components as affected, including Conversations, Codex, Responses and the Agents API, before reporting full recovery.

Warp Factories can now be used from Microsoft Teams and Azure DevOps.

Reception is an AI receptionist that answers calls and books jobs; Architect drafts improvements and tests them in simulated conversations.

Chats, projects and settings moved automatically, and every account can now invite collaborators.

Balanced, the default, uses about 70% fewer credits than Flagship, which needs Pro+ or Max.

Remote Control is on by default except for Enterprise orgs; agents keep running locally.

OpenAI added a visual ad format and measurement tools in October, after expanding ChatGPT Ads to Southeast Asia and Taiwan (60+ countries) in September.

Qodo pitches 3.0 as the quality and governance layer for an agentic software factory.

The plugin connects agents and subscriptions you already use, ships with no agents installed, and is in 2026.3 EAP builds.

The setting is tied to the plan, members can't re-enable it, and Enterprise admins can audit access to shared notes.

Draft mode is suggested automatically on all plans.

Sonnet 5.5 is in Power Mode and GPT-6.1 Sol in Max Mode; availability depends on tier and workspace policy.

Experimental support for Pro and above, with a 1M-token context and adaptive thinking on by default.

Several chats can now work on the same branch at once, each keeping its own context.

Reusable multi-step agent plans arrived in the IDE, the CLI and (opt-in) Kiro Web cloud sessions.

Google aims Argon at real-world coding, enterprise knowledge work and cyber defense, and said it would roll out soon.

SuperGrok subscribers can use Grok in the Agent Panel, and subagents can pick a model per spawn.

The beta is now open to everyone after a month of limited testing.

Two new creation tools for making videos and games inside Manus.

The heygen-video-1 endpoints generate 5–15 second clips at 480p or 768p from prompts of up to 5,000 characters.

Skills are reusable instructions Gemini can build from chats and run on matching prompts; existing Gems will be migrated automatically when Gems go away.

Hold a key and speak, and clean text appears wherever your cursor is.

Impact ran from 7:00 to 7:59 am PT on 29 Sep; sign-in, new chats and Code sessions failed, and some messages sent then may not have been saved.

OpenAI's recap gathers its DevDay launches, covering GPT-6 Astra, ChatGPT, Codex, the APIs, security and builder tools.

Muse can now generate on-brand designs straight from a user's Canva account.

v4 brings more expressive speech and better voice cloning in 90+ languages; v4 Turbo targets real-time agents at about 100 ms median latency.

Dots are assistants that carry on with complex projects and everyday tasks while you stay in control.

Plaid uses OpenAI's ultrafast tier for GPT-6 Astra modes and works only with Amp-provided inference, not linked ChatGPT subscriptions.

Manus says Cascade used 23% fewer tokens and cost 32% less in one tested configuration; Automations can trigger on events in connected services.

Sonnet 5.5 is the second model in the Claude 5.5 family, positioned as a faster, lower-cost complement to Opus 5.5.

Each Team Bot combines shared context, plugins, credentials and memories, keeps each person's chats private, and can join Slack channels.

Opus 5.5 replaces GPT-5.6 Sol in medium mode; Amp says it solved 65% of internal eval tasks for 10% less than Sol, while ChatGPT Only users stay pinned to Sol.

Templates built in Flows can now run via the ElevenLabs API.

The Agent and Orchestrator APIs drop gpt-5.4 variants, gpt-5.2, gpt-5.1, gpt-5 and gpt-5-mini at 00:00 UTC.

Fast models preview your prompt across styles, defaults can be saved, and edit-model inpainting now changes only the selected pixels.

Google adds a near real-time visual avatar to its conversational Live model.

A new developer portal handles submissions and review tracking, and shows usage analytics once a plugin is live.

Google calls them its most expressive audio models yet.

A new wave of Connected Apps is rolling out to the Gemini app.

A failed database maintenance caused API errors from 12:57 am PDT on 23 Sep, with Projects backlogs fully cleared at 9:55 pm PDT.

Rollouts monitors each change as it deploys and reports health per environment, and Security Review reports exploitable bugs on every pull request.

Two GPT-6 models for everyday work that trade capability against cost differently.

Paid plans get a DaVinci Resolve plugin and four Kling video models in Runway MCP.

Lovable's benchmarks show the same build quality as Opus 5 with fewer steps.

Both models stopped generating on 24 Sep; flows using a Sora node need another video model, and Seedance 1.5 Pro retires on 11 Nov.

The model is fine-tuned for clinical audio.

Users can browse models, review token pricing and set Thinking or Fast Mode; Enterprise teams must opt in to custom models.

Each model is grounded in company knowledge and brand guardrails through Jasper IQ.

xAI reported outages across the API, grok.com and Grok Build from about 6:03 to 6:28 pm PDT on 21 Sep.

Edit, caption and export without leaving the project.

Founder Tomas Halgas and three engineers join to make it easier to understand how software works.

Embed a lip-synced avatar with turn-taking, and read credit balance for the billing period programmatically.

Custom connectors (beta) take an endpoint and auth, and enterprise audit logs gain 65+ event types.

xAI pitches it as more accurate and cheaper to run.

Canto is rolled out to everyone; Wispr says it had the lowest word error rate on its real-world English test set.

Gen media models sit underneath, with focused creative apps on top.

The four professional tools now come together as one connected suite for professional creatives.

It works up to 4K and 5 minutes long.

Everything Cowork does now works from any conversation without picking a mode, and designs, decks and docs can be made in any chat.

The map describes the system as it exists today.

Notes are written in the background as you work and read back when you return.

Google calls them its most advanced live dialogue models yet.

The bot-free meeting notetaker that launched on Mac in August now works on Windows too.

Tabnine CLI, powered by OpenCode, reached all customers with one installer, plus an admin-enabled agent beta in VS Code and JetBrains.

Skills are shared, reusable instructions that can run automatically or be downloaded as SKILL.md files for Claude Code, Codex, Cursor and others.

The plugin brings accounts, opportunities and pipeline into Claude on all paid plans, for organizations Salesforce approves.

A new desktop app built to work alongside your other tools and apps.

Aimed at research, modeling and client-ready materials.

Projects run in the cloud by default, keep shared context over months of work, and can follow Slack channels, schedules or PRs without being prompted.

A managed service for building cloud agents, with orchestration, long-running sessions and tool use.

The new image model turns ideas, sketches and reference photos into more personalized, polished images.

The Series D round values the company at more than €21 billion post-money.

Google says its best-sounding music model brings more expressive vocals.

It supports instruction edits, up to 4 image references, inpainting and outpainting.