What shipped · 2026-w44
What shipped in AI dev tools: week of Oct 26, 2026
Oct 26, 2026 – Nov 1, 2026
83 releases across 17 tools this week; biggest: GitHub Copilot (24 releases).
GitHub Copilot
Agent Merge on GitHub pull requests
On October 28, 2026, Agent Merge was added to pull requests on GitHub. It was already available in the GitHub Copilot app and Visual Studio Code. Select Watch and fix in the pull request merge box and choose whether Copilot fixes failing Actions workflows, resolves merge conflicts, or both. When a selected blocker appears, Copilot works on it in a cloud development environment. Agent Merge can be enabled by default for every pull request you open, and it can be combined with auto-merge so GitHub merges the pull request once requirements are met. It is available on paid Copilot plans, subject to organization and enterprise policies.
Source for Agent Merge on GitHub pull requests (opens in a new tab)Cloud Copilot Automations in VS Code
On October 28, 2026, Copilot Automations in Visual Studio Code can run in the cloud. Personal scheduled automations keep running when Visual Studio Code is closed. Personal automations can also run on GitHub events, such as an issue being assigned or a pull request being opened, and shared cloud automations can be created for the repository. Update to Visual Studio Code 1.144 and enable chat.automations.cloud.enabled. Automations are in the Automations tab of the Agents window. Copilot Automations is available on Copilot Pro, Pro+, Max, Business, and Enterprise.
Source for Cloud Copilot Automations in VS Code (opens in a new tab)Connected Copilot cloud sessions
On October 28, 2026, GitHub made a connected Copilot cloud experience generally available on paid Copilot plans, rolling out starting October 28. A session can start on github.com, GitHub Mobile, Copilot CLI, the Copilot app, VS Code, or from Slack or Microsoft Teams, and continue on another of those surfaces with the same session list. By default, every session runs in a cloud sandbox on GitHub-managed infrastructure. Sandboxes have no one-hour limit on total running time, and a snapshot is saved when a sandbox shuts down so the session can resume later. Billing for cloud sandbox usage begins in November 2026. The included Actions entitlement is now one dollar-based pool for standard GitHub-hosted runners, cloud sandboxes, and, on Team and Enterprise Cloud, larger hosted runners, instead of a fixed number of standard runner minutes. Each product keeps its own published rate. Sessions that use Actions-based self-hosted runners or copilot-setup-steps.yml still need Actions as the environment. Running Copilot CLI with --environment-host registers your own computer or VM. Cloud sessions include Bash, the GitHub CLI, code search, connectors, custom MCP servers that use key-based authentication, skills, and slash commands such as /rubber-duck and /context on github.com and GitHub Mobile. On GitHub Mobile, a task can be described by voice or with an attached image or file. Access is governed by the Copilot in the cloud policy, and a repository previously opted out of the Copilot cloud agent stays opted out. Without write access to a repository, Copilot there is read-only. Copilot Free can keep using Copilot Chat on github.com and GitHub Mobile.
Source for Connected Copilot cloud sessions (opens in a new tab)Copilot app projects and computer use
On October 28, 2026, the GitHub Copilot app added Projects in public preview. A project holds a goal plus context such as repositories, design mocks, Slack threads, or notes, and the models and tools for the job. Copilot runs as many agents as the work needs and brings back results, such as pull requests. Project automations can repeat a task, such as creating a promo video when a feature lands. A single-repository task can still start as a session from that repository. The app can use Slack and Figma context to research a request, draft a specification, or work through a design, and it can generate and edit images in chat. Computer use, in public preview on macOS and Windows, lets Copilot read app content, click controls, enter or edit text, and navigate across apps, including apps without an API, command-line interface, or MCP integration. Copilot asks for approval before controlling an app, and an organization can disable the feature in managed settings. Turn it on under Settings, Computer Use. On macOS, grant Accessibility and Screen Recording permissions.
Source for Copilot app projects and computer use (opens in a new tab)Copilot Automations can propose their own improvements
On October 28, 2026, Copilot Automations can learn from outcomes on GitHub and propose improvements to themselves. Examples GitHub gives are whether people change labels an automation applied, whether developers accept its review comments, and whether maintainers merge the pull requests it opens. A proposed change to the prompt, tools, or triggers is sent for review before it is applied. Asking Copilot to set up a self-improving automation can create a group, such as a review automation on new pull requests and a weekly automation that proposes refinements to it. Copilot Automations is available on Copilot Pro, Pro+, Max, Business, and Enterprise.
Source for Copilot Automations can propose their own improvements (opens in a new tab)Copilot Automations triggers, tools, and credit limits
On October 28, 2026, Copilot Automations added triggers, tools, and AI credit limits. The Trigger an automation REST API can start an automation from an external system. Twenty-four more GitHub events can start one, across issues, pull requests, discussions, and code scanning, secret scanning, and Dependabot alerts. A trigger's "Can be triggered by" setting defaults to Only users and can be set to Users and Copilot so another automation's event fires it. Automation run succeeded and Automation run failed triggers can chain automations. Twenty new GitHub tools include uploading image or video attachments, retrying workflow runs, merging pull requests, updating comments, marking a pull request ready or converting it to a draft, hiding comments and reviews, removing reactions, and reading or commenting on discussions. An automation can also use selected tools from a custom MCP server added in the repository's settings. Optional credit limits can be set per run, day, week, or month. A per-run limit stops that run. A daily, weekly, or monthly limit blocks new runs until the period resets. Automations without limits are unchanged. Copilot Automations is available on Copilot Pro, Pro+, Max, Business, and Enterprise, in repositories where you have write access.
Source for Copilot Automations triggers, tools, and credit limits (opens in a new tab)Copilot code review Max and approvals
On October 28, 2026, Copilot code review added Max, in public preview, and made Copilot approvals generally available for Copilot Pro, Pro+, Max, Business, and Enterprise. Max joins Lite and Balanced. It runs several independent review agents across models from different providers, then an aggregation agent reconciles their findings. Max takes longer and uses more AI credits than Lite or Balanced. It can be selected for one pull request or set as the default for a repository, organization, or enterprise, and each level can override the one above it. Administrators can set a maximum effort level so higher levels are unavailable. Approvals are off by default. When they are on, every review includes an approval assessment, and Copilot's approval counts toward required approvals. New commits dismiss that approval. Administrators turn approvals on at the enterprise, organization, or repository level, and repository administrators can limit which file paths Copilot may approve.
Source for Copilot code review Max and approvals (opens in a new tab)Copilot Fusion can route to local models
On October 28, 2026, GitHub Copilot Fusion added experimental routing between supported on-device models and cloud models in Copilot CLI, the Copilot app, and Visual Studio Code. On supported Windows devices, MAI Code 1.1 Flash can be downloaded through the Windows ML provider. In the CLI, enable experimental mode and follow the prompts in /model. In the app, configure Windows ML under Settings, Model providers, enable Local model auto routing under Settings, Experiments, and start a new session. Windows ML local models are enabled by default on Windows. In VS Code, enable chat.copilot.localModel.enabled. Select Standard in Copilot Fusion to let Copilot choose between local and cloud models, or select the local model directly. On Windows, macOS, and Linux, Copilot CLI can also route to a manually configured MAI Code 1.1 Flash or Qwen3.8 Flash Next served by a local OpenAI-compatible server. Discovering a model through bring-your-own-key configuration does not by itself make it eligible for Fusion. The Copilot app can select compatible local models through bring-your-own-key, but its intelligent routing currently supports managed Windows ML models only. Routing decisions still happen in the cloud, so local inference does not make the session offline.
Source for Copilot Fusion can route to local models (opens in a new tab)Copilot in JetBrains, Universe releases
On October 28, 2026, GitHub Copilot in JetBrains IDEs added Copilot Fusion, with Standard, High, and Max effort levels. Fusion replaces the Auto model option and is rolling out gradually. GitHub says Fusion use is discounted 30% for the first 30 days. Local sandboxing is generally available, so tools and commands Copilot starts run with restricted access to the filesystem, network, credentials, and other system capabilities. Next edit suggestions are generally available. The Microsoft Copilot agent harness is the default when it is available. A session sidebar can open sessions side by side, and settings can choose the default permission mode for new sessions. The update also improves edit reliability, cancellation recovery, and selected-code context.
Source for Copilot in JetBrains, Universe releases (opens in a new tab)Copilot in VS Code, Universe releases
On October 28, 2026, GitHub Copilot in Visual Studio Code added ways to keep one agent session across clients and machines. A cloud agent session started from the Agents window keeps running on GitHub's infrastructure after the laptop closes, and the same session can continue in VS Code, the Copilot app, GitHub Mobile, github.com, or Copilot CLI, including sessions started in a Slack or Teams direct message. Agent Host isolates sessions from individual VS Code windows so work continues after the project window closes. Remote SSH and dev tunnels can run that host on another machine, which has to stay running. The editor and the Agents window stay on the same live session. From a project folder, code agent host --github-environment lets GitHub Mobile steer the agent on that machine. The Copilot harness, built on the Microsoft Copilot SDK, is now the default for new sessions. VS Code can install Copilot CLI through its Windows installer, when adding VS Code to the path on macOS, or by running copilot on Linux. The Agents window can run models and harnesses in parallel, ask a Judge model to recommend one, and use Synthesize to combine attempts. That comparison is in preview. Computer use is experimental: enable chat.agentHost.computerUse.enabled, restart, and run /computer on. Each application requires approval. The Agent Host Protocol lets another client connect to code agent host, with libraries for TypeScript, Rust, Go, Kotlin, and Swift. Customizations in the Agents window browse skills, MCP servers, and plugins. The Customization Marketplace is in preview. Administrators can require or block named plugins and marketplaces, and .github/copilot/settings.json can register default plugin marketplaces without enforcing them. Copilot connectors, in preview on all plans, connect a service such as Slack, Notion, or Microsoft Learn once and reuse it across Copilot clients on the Copilot harness. Business and Enterprise administrators need to allow Editor preview features for the marketplace and connectors.
Source for Copilot in VS Code, Universe releases (opens in a new tab)Copilot included with GitHub Enterprise
On October 28, 2026, GitHub said enterprises that work with a GitHub or Microsoft account team can get Copilot included with GitHub Enterprise for every licensed user. AI usage is metered from the first request against the enterprise budget, with no included usage allowance. Code completion and next edit suggestions keep working when no budget is available. Copilot Chat, Copilot code review, and Copilot in the cloud require available budget. An exhausted budget does not remove the seat. Users in the enterprise account who do not hold a GitHub Enterprise license get the same access through Copilot Standalone, at $10 per user per month. That access has the same features and models as Copilot Enterprise and does not include monthly AI credits per seat. It is available now for GitHub Enterprise Cloud customers who work with an account team. Self-serve availability is opening in stages, and current self-serve subscriptions are unaffected. Copilot Business and Copilot Enterprise remain available to purchase.
Source for Copilot included with GitHub Enterprise (opens in a new tab)Copilot SDK is becoming Microsoft Copilot SDK
On October 28, 2026, GitHub said the GitHub Copilot SDK is becoming the Microsoft Copilot SDK, the programmable interface for the Microsoft Copilot harness used across Microsoft and GitHub products. The SDK repository, branding, and documentation are being updated in phases. Existing package names, namespaces, and some documentation may still say GitHub Copilot SDK during the transition. GitHub says no immediate changes are required for existing users, and support for the SDK stays the same.
Source for Copilot SDK is becoming Microsoft Copilot SDK (opens in a new tab)GitHub Agentic Workflows generally available
On October 28, 2026, GitHub Agentic Workflows became generally available. Markdown workflow files compile into GitHub Actions YAML and run a coding agent for tasks such as issue triage, CI failure analysis, documentation updates, and test writing. The default engine is GitHub Copilot CLI, which requires a Copilot license, and a workflow can bring its own coding agent. Agents follow integrity filter rules for GitHub content, run with read-only permissions by default, and execute in a sandboxed container behind the Agent Workflow Firewall. Outputs go through the safe outputs process, and a threat detection job scans proposed changes before they are applied. Since the public preview, workflows can run on Actions Runner Controller runners with Docker-in-Docker, gh aw forecast includes threat detection credits and monthly ranges, a model provider can be chosen per workflow, a model router can optimize for task type and difficulty, and agents can run in a Cloud Hypervisor or NVX micro-VM. It is generally available on github.com for GitHub Free, Team, and Enterprise.
Source for GitHub Agentic Workflows generally available (opens in a new tab)GitHub Copilot Fusion generally available
On October 28, 2026, GitHub Copilot Fusion became generally available. Selecting it does not pick one model. It routes and orchestrates models for each task, combining Project HydraFusion's multi-model orchestration with Copilot auto model selection. Effort levels are Fusion Standard, Fusion High, and Fusion Max. Workflows are Single (one model), Cascade (an efficient model drafts, then a quality check can escalate), and Critique (one model drafts, another reviews, and the first revises). GitHub says Copilot Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, at about half the average API cost per task, using about half as many tokens and taking about 40% less time per task. Fusion uses only models enabled in enterprise and organization policies. Enterprise administrators can disable a model for Fusion without disabling it for everyone, and can set Fusion, at any effort level, as the default. Each model Fusion uses is billed at that model's standard rate. For a limited time, paid Copilot plans get a 30% discount on model costs when using Fusion. It is available on all Copilot plans in Visual Studio Code, the GitHub Copilot app, GitHub Copilot CLI, and JetBrains IDEs. Previous auto, HydraFusion, or HydraFusion Max selections move to the corresponding Fusion effort levels.
Source for GitHub Copilot Fusion generally available (opens in a new tab)GitHub intelligence: memory, dreaming, and code search
On October 28, 2026, GitHub introduced GitHub intelligence for agents, with memory, dreaming, and code search. GitHub Copilot Memory is generally available on paid Copilot plans. Memories of personal preferences and repository conventions can be created in the Copilot app, Copilot CLI, VS Code, the Copilot cloud agent, Copilot code review, Slack, Teams, and GitHub Mobile. Sessions can generate and curate memories. GitHub says that among reviewed Copilot CLI public-preview sessions where the agent used memory, 93% showed better adherence to developer preferences and 42% showed more efficient workflows. Personal memories can be viewed or deleted in Copilot settings. Copilot dreaming, in public preview, is opt-in and consumes AI credits. Dream types are personal memory curation, repository customization (including proposed pull requests that update skills and custom instructions), and code review customization that turns recurring accepted review feedback into guidance. Individuals enable personal curation in Copilot settings and repository dreams per repository. Organization and enterprise administrators enable dreaming through Copilot policies. GitHub code search lets Copilot look across related repositories for exact text, symbols, regular expressions, and semantic matches, including through /research.
Source for GitHub intelligence: memory, dreaming, and code search (opens in a new tab)GitHub-verified Copilot connectors
On October 28, 2026, GitHub-verified MCP servers, called connectors, started rolling out in public preview. Connect a service once, with no local server to install, and use it in the GitHub Copilot app, Copilot CLI, VS Code, Copilot in the cloud, and GitHub Mobile. At launch, 26 services can be connected, including Slack, Figma, Notion, Sentry, Linear, Atlassian Rovo, Cloudflare, Vercel, and Work IQ. In the Copilot app, open Customize and select Connect. In Copilot CLI, run /mcp and use the Connectors tab. In VS Code, connect from Customizations in the Agents window. On github.com, use the Copilot customization page. Copilot uses a connected service's tools when they are relevant to the request, and the Copilot app can use a connected service in scheduled automations. The rollout covers all Copilot plans. Copilot Business and Copilot Enterprise need the Editor preview features policy. If an enterprise MCP allowlist is set in managed settings, an administrator also has to add GitHub-verified MCP servers to it. Some services also need an administrator to allow access on that service's side.
Source for GitHub-verified Copilot connectors (opens in a new tab)Image generation in Copilot chat
On October 28, 2026, GitHub Copilot chat can generate and edit images from natural language. The selected reasoning model calls an image-generation tool powered by GPT Image 2.5 Flare, previews the image inline, and saves it to the project. It works in VS Code, the GitHub Copilot app, and Copilot CLI. Organization and enterprise owners can turn it off by disabling the image generation model in Copilot model policies. A visible "Made with AI" watermark is opt-in and off by default. Image generation is in public preview for Copilot Pro, Pro+, Business, and Enterprise. It is billed separately from the chat model, based on image tokens that count toward the plan's AI credits.
Source for Image generation in Copilot chat (opens in a new tab)Security extension for the Copilot app
On October 28, 2026, the GitHub security extension for the GitHub Copilot app entered public preview. Where code scanning is already enabled, it lists open code scanning alerts by severity, with search and filters, and can open an alert's details and affected code. Selecting one or more alerts and asking Copilot to Explain or Fix changes only the local working copy. It shows linked pull requests, and dismissing an alert is sent for approval when the organization uses delegated alert dismissal. /security-review checks the current changes and says when the code has changed since the last review. Install it with copilot plugin install github-security@awesome-copilot, or from Customize in the Copilot app. Explain, Fix, and /security-review use the selected model and consume AI credits. Browsing and dismissing alerts do not. Any Copilot plan can install the extension and run /security-review. Seeing, fixing, and dismissing existing alerts requires permission to view them. Private repositories need GitHub Code Security. Copilot Business and Copilot Enterprise need the GitHub Copilot app policy enabled. It runs on github.com and GitHub Enterprise Cloud with data residency, not on GitHub Enterprise Server.
Source for Security extension for the Copilot app (opens in a new tab)Shared Copilot Automations
On October 28, 2026, shared Copilot Automations started rolling out. A shared automation is visible to everyone with read access to the repository and editable by anyone with write access. AI credits are billed to the organization or user that owns the repository, and actions on GitHub are attributed to Copilot rather than to one person. The automation stays with the repository if its creator leaves, and change history records how the tools, triggers, and prompt evolved. When creating an automation on github.com, the GitHub Copilot app, or Visual Studio Code, choose Shared under Type. The type cannot be changed later. An organization-owned repository also needs the policy that allows use of the Copilot cloud agent billed to the organization. The rollout covers Copilot Pro, Pro+, Max, Business, and Enterprise, and availability may vary.
Source for Shared Copilot Automations (opens in a new tab)Copilot server-to-server authentication
On October 27, 2026, GitHub Copilot added server-to-server authentication for GitHub Apps installed on organizations, in public preview. Services and agents built with the Copilot SDK can make requests for an organization with a short-lived installation access token, without a signed-in user's credentials. The app needs the read-only Copilot Requests organization permission and must be installed on the organization. Usage is attributed and billed to the organization that owns the installation, and that organization's model policies, budgets, and spending controls apply. Organization owners can see which apps have access, find app-attributed usage in AI usage CSV reports in the app_client_id field, and stop an app by suspending or uninstalling it.
Source for Copilot server-to-server authentication (opens in a new tab)GitHub Copilot CLI 1.0.104
Released October 27, 2026. The managed model fusion starts sessions on Auto and applies fusionTier settings. Interactive and SDK Autopilot sessions keep waiting for attached background tasks instead of stopping after COPILOT_TASK_WAIT_TIMEOUT_SECONDS; -p --autopilot still honors that timeout. When a saved Auto preference is no longer available, Auto falls back to its default preference instead of failing each turn. Assisted Permissions are available without enabling experimental mode.
Source for GitHub Copilot CLI 1.0.104 (opens in a new tab)GitHub Copilot CLI 1.0.105
Released October 27, 2026. Selecting Copilot Fusion High or Max in /model now works from long-context sessions.
Source for GitHub Copilot CLI 1.0.105 (opens in a new tab)Copilot assisted permissions generally available
On October 26, 2026, assisted permissions became generally available in GitHub Copilot CLI, the GitHub Copilot app, and Visual Studio Code. When Copilot requests permission to run a command, access a file system path, or open a URL, a large language model evaluates the request. Requests classified as clearly safe proceed automatically. Other requests require approval. In the CLI, run /permissions and select assisted. In the Copilot app, open the mode picker, click Tool permissions, and select Assisted. In Visual Studio Code, open the permissions picker and select Assisted permissions. Enterprise admins can block Allow All while keeping Assisted available, or disable Assisted separately, and can choose the default permission mode for new sessions. If that default is unavailable or prohibited by policy, the session starts with manual approval.
Source for Copilot assisted permissions generally available (opens in a new tab)GPT-6.1 Terra in GitHub Copilot
On October 26, 2026, GPT-6.1 Terra became generally available in GitHub Copilot. GitHub says it is designed for everyday agentic coding and terminal work at a lower cost than GPT-6.1 Sol, and that in early testing it solved noticeably more terminal and C# coding tasks than GPT-5.6 Terra. It is available to Copilot Pro, Pro+, Max, Business, and Enterprise users in the model picker in Visual Studio Code, Visual Studio, Copilot CLI, the GitHub Copilot cloud agent, the GitHub Copilot app, github.com, GitHub Mobile on iOS and Android, JetBrains IDEs, Xcode, and Eclipse. Rollout is gradual. It is billed at provider list pricing under usage-based billing. Copilot Enterprise and Copilot Business admins manage access through the model policy. Under default model enablement, new models are enabled unless an admin has turned off the global default or disables this model.
Source for GPT-6.1 Terra in GitHub Copilot (opens in a new tab)
Vercel
Daemon liveness checks
On October 30, 2026, Vercel Daemons added HTTP liveness checks. A livenessProbe in the daemon configuration points at an HTTP endpoint that returns an error status when the daemon is no longer healthy. Vercel periodically requests the endpoint and restarts the daemon's container when the check fails.
Source for Daemon liveness checks (opens in a new tab)Daemon service bindings
On October 30, 2026, Vercel Daemons could call the services in their deployment through service bindings. A binding on the daemon names the target service and the environment variable that receives the service's URL. The daemon calls that URL like any other HTTP endpoint. Vercel handles the internal routing, authentication, and TLS, and the service stays private. Bindings point one way: services cannot bind to daemons. Each instance keeps calling the deployment that started it, even after a promotion or rollback.
Source for Daemon service bindings (opens in a new tab)Data Cache moves to Runtime Cache
On October 30, 2026, Vercel Data Cache started moving to Runtime Cache. Next.js fetch caching and unstable_cache switch over on every version without code changes. On Pro, each project and environment gets its own Runtime Cache storage allocation, with 8x the space of Data Cache's previous team-wide pool. Enterprise projects get 10x their previous Data Cache storage. After a team migrates, Data Cache reads and writes that were previously unbilled are charged at Runtime Cache pricing. Rates start at $0.40 per million read units and $4.00 per million write units, and vary by region. Hobby teams get 1 million read units and 200,000 write units free each month across all projects. On Pro, usage draws from the $20 monthly infrastructure credit, and anything beyond it is billed at regional rates. Teams created on or after October 30, 2026 use Runtime Cache by default. Existing Hobby teams migrate on November 29, 2026. Vercel notifies Pro teams of their migration date. Enterprise account teams contact those customers with migration details. Existing Hobby and Pro teams can opt in early from the team's CDN settings, and Runtime Cache pricing applies as soon as they opt in.
Source for Data Cache moves to Runtime Cache (opens in a new tab)Finer IP geolocation
On October 29, 2026, Vercel IP geolocation started resolving visitor locations with finer network granularity. Vercel says it delivers city-level data for nearly all requests and postal-code data for 90% of traffic, and that the data is refreshed more often. It is on for every project with no code or configuration change. Existing geolocation headers and the geolocation() helper in @vercel/functions return the same fields for Vercel Functions and Routing Middleware. Subnets announced from a hosting provider in a different country than their regional internet registry record now resolve to the data center where the traffic terminates. Some visitors can resolve to a different city or country, most often traffic from VPNs, hosting providers, and leased IP ranges.
Source for Finer IP geolocation (opens in a new tab)Node.js 26 LTS on Vercel
On October 29, 2026, Node.js 26 LTS became generally available for builds and Vercel Functions, and it is the default for new projects. Built-in fetch uses HTTP/2 by default through Undici 8. The release includes the Temporal API and V8 14.6, which adds Map.prototype.getOrInsert(). If engines.node in an existing project's package.json allows 26, such as >=20, the next deployment uses Node.js 26. Set engines.node to 26.x to choose it directly. Prebuilt deployments need Vercel CLI 65.8.0 or later. Set VERCEL_FETCH_DISABLE_H2 to 1 and redeploy to keep HTTP/1.1 in Functions. Minor and patch updates to Node.js 26 roll out automatically.
Source for Node.js 26 LTS on Vercel (opens in a new tab)Separate AI crawler and assistant firewall rules
On October 29, 2026, the AI Bots managed ruleset in Vercel Firewall could log or block AI crawlers and AI assistants separately. Vercel describes AI crawlers as bots that collect content for indexing and model training, and AI assistants as bots that fetch pages in response to user actions or search queries. Vercel updates the rules as new bots are added. Configure it under the project's Security, Firewall, Rules, then Bot Management. vercel firewall diff reviews the changes and vercel firewall publish applies them. The ruleset is available on all plans.
Source for Separate AI crawler and assistant firewall rules (opens in a new tab)Vercel Connect webhook triggers
On October 29, 2026, Vercel Connect added webhook triggers for Resend, X, Twilio, Notion, and Telegram. Connect receives events, such as Telegram bot messages or emails received by a Resend domain, verifies the requests, handles provider heartbeats and verification challenges, and forwards them to an endpoint in a Vercel project. Connect registers webhooks automatically for Resend and Telegram. For Twilio, Notion, and X, copy the webhook URL from Connect and finish setup in the provider's settings. These connectors join existing trigger support for Slack, Discord, Microsoft Teams, GitHub, and Linear.
Source for Vercel Connect webhook triggers (opens in a new tab)Vercel Plugin in Antigravity
On October 29, 2026, the Vercel Plugin became available in the Google Antigravity marketplace. It gives an Antigravity agent guidance on Vercel APIs and recommended patterns, with skills for products such as the AI SDK, Vercel Functions, and Workflow. Install it from Customizations, Marketplace in the Antigravity app.
Source for Vercel Plugin in Antigravity (opens in a new tab)Agents can enable Web Analytics from the CLI
On October 28, 2026, Vercel CLI stopped requiring an interactive terminal to enable Web Analytics or Observability Plus, so an agent can turn them on. From a linked project directory, run vercel project web-analytics enable or vercel project observability enable. Enabling Observability Plus requires the team to already be subscribed to it. This requires Vercel CLI 64.2.0 or later.
Source for Agents can enable Web Analytics from the CLI (opens in a new tab)Box Mount in Vercel Sandbox
On October 28, 2026, Box Mount became available in private preview for Vercel Sandbox. It syncs a Box folder with a directory in the sandbox so an agent can read and write those files through the filesystem. Unmount before stopping the sandbox, and the agent's writes are in Box after the sandbox is deleted. The sync process and the agent run as separate sandbox users, so the Box credentials stay out of the agent's workspace. Request access from Box, then follow Vercel's Box Mount documentation.
Source for Box Mount in Vercel Sandbox (opens in a new tab)Paseo agents on Vercel Sandbox
On October 28, 2026, Paseo could run fx, Codex, Claude Code, OpenCode, and Pi in Vercel Sandbox through the Vercel Sandbox plugin. An agent started from Paseo on a laptop can be controlled from a phone after the laptop is closed. Stopping a sandbox saves files and conversation history. Active processes stop with the sandbox. The Vercel AI Gateway key is entered once during setup and reused for model requests, outside the sandbox. Enable plugins in Paseo under Settings, Plugins, then run paseo plugin add vercel-labs/paseo-vercel-sandbox.
Source for Paseo agents on Vercel Sandbox (opens in a new tab)Vercel Agent sessions use fewer tokens
On October 27, 2026, Vercel said Agent sessions in the dashboard and Slack cost 67% less on average. Agent uses fewer tokens per session and reuses more of the prompt cache, including in long Slack threads. Updated Vercel CLI skills give agents clearer guidance for common workflows and error recovery, and other agents that use those skills get the same guidance. Billing rates are unchanged: paid sessions are billed at the model provider's inference rates plus a Vercel Agent fee of $0.25 per million tokens. The change is already live.
Source for Vercel Agent sessions use fewer tokens (opens in a new tab)Vercel Plugin in OpenCode
On October 27, 2026, the Vercel Plugin became available in OpenCode 2. It gives on-demand guidance on Vercel APIs and recommended patterns, with skills for the AI SDK, Vercel Functions, Workflow, and more, plus specialist agents for deployment, performance optimization, and AI application design. /vercel:deploy deploys a project and /vercel:env manages environment variables. Install it with opencode plugin add github:vercel/vercel-plugin, then run /mcps, select vercel, and sign in to connect the Vercel MCP server.
Source for Vercel Plugin in OpenCode (opens in a new tab)Vercel Agent can validate alerts before paging
On October 26, 2026, Vercel Agent can investigate alerts from Vercel monitors before they reach a paging system. With validation enabled for a webhook destination, only alerts Agent classifies as Critical are sent to it. Turn it on from a monitor's notification settings by selecting Critical Only on that destination, or with the Vercel CLI command vercel monitors notifications and --critical-only. Agent checks logs and metrics, and a monitor can include context to guide that check. Alert validation is available on Pro and Enterprise plans with Observability Plus, which includes 10 Agent investigations per billing cycle. Additional investigations are billed separately.
Source for Vercel Agent can validate alerts before paging (opens in a new tab)
Vercel AI SDK
Ling 3.1 Flash 1019 on AI Gateway
On October 30, 2026, Ling 3.1 Flash 1019 from InclusionAI became available on Vercel AI Gateway, free until November 3, 2026. Vercel says the model is designed for coding, multi-step reasoning, and agents that use tools, with a 1M-token context window. The model ID is inclusionai/ling-3.1-flash-1019. Free requests still appear in the spend dashboard with a request trace. On November 3, requests begin charging at the published rate. It can be selected in Claude Code, Codex, Cursor, and other coding agents after npx vercel ai-gateway setup.
Source for Ling 3.1 Flash 1019 on AI Gateway (opens in a new tab)Mistral Large 4 updated weights on AI Gateway
On October 29, 2026, Mistral Large 4 on Vercel AI Gateway started using updated weights, with lower pricing for input, cached input, and output tokens. Vercel says Mistral's announcement reports improvements in agentic coding and defensive cybersecurity, greater token efficiency on coding tasks, spreadsheet creation and editing, web research, and resistance to prompt injection. The model ID stays mistral/mistral-large-4.
Source for Mistral Large 4 updated weights on AI Gateway (opens in a new tab)Strix on Vercel AI Gateway
On October 29, 2026, Strix, the open-source AI penetration testing tool, could run on Vercel AI Gateway. Vercel says that covers any of the 400+ models on the Gateway, including models from Anthropic, OpenAI, Kimi, GLM, and Deepseek, through one API with no markup, and that usage and cost are tracked in AI Gateway. Set STRIX_LLM to vercel_ai_gateway/ followed by the standard provider/model ID, and set LLM_API_KEY to an AI Gateway API key.
Source for Strix on Vercel AI Gateway (opens in a new tab)fx in T3 Code
On October 28, 2026, fx, Vercel's open-source coding agent harness, became available in T3 Code on web, desktop, and mobile. T3 Code can add agents as providers for coding sessions. The built-in fx provider supports Supervised and Auto modes, along with the models and skills already used with fx. Install fx, set up a model provider, then add fx in T3 Code from Settings, Providers, Add provider.
Source for fx in T3 Code (opens in a new tab)Particle-1 on AI Gateway
On October 28, 2026, Particle-1 from Particle AI became available on Vercel AI Gateway as particle/particle-1. Vercel says it is post-trained on DeepSeek V4.1 Flash to complete software engineering tasks with fewer tokens. Particle describes the gains as shorter reasoning, fewer rounds of trial and repair, and more focused exploration. It has a 1M-token context window. Call it through the AI SDK, fx, the OpenAI-compatible Chat Completions API, or a coding agent after npx vercel ai-gateway setup. AI Gateway reflects provider pricing with no markup and does not charge a platform fee on inference, including Bring Your Own Key requests.
Source for Particle-1 on AI Gateway (opens in a new tab)Virtual models on AI Gateway
On October 28, 2026, virtual models became available on Vercel AI Gateway. A virtual model is a saved model and settings configuration, called with a custom ID such as vmc/coding-agent. Changing the underlying model or settings in AI Gateway does not require redeploying apps or reconfiguring agents. It can set reasoning effort, fast mode, and prompt caching, choose providers and fallback models, require Zero Data Retention, and attach custom reporting tags. Create one on the Virtual Models page or with vercel ai-gateway virtual-models create. Replace the provider model ID with vmc/ and the slug in the AI SDK or a compatible API. For a coding agent, connect with vercel ai-gateway setup and select the virtual model. Virtual models work with every model type on AI Gateway.
Source for Virtual models on AI Gateway (opens in a new tab)fx harness adapter runs on the host
On October 27, 2026, the fx adapter for the AI SDK's HarnessAgent started running the harness and model requests in the same process as the application. Built-in filesystem and shell operations stay in the sandbox. The adapter uses libfx and no longer installs fx inside the sandbox, runs a sandbox bridge, or exposes a sandbox port, so model credentials stay on the host. It adds built-in tool filtering, mid-turn steering, structured output that matches a schema, and image tool results through toModelOutput(). Sessions can be saved and resumed in a new host process. Update @ai-sdk/harness-fx to use it.
Source for fx harness adapter runs on the host (opens in a new tab)HarnessAgent accepts image inputs
On October 27, 2026, the AI SDK HarnessAgent layer started accepting image inputs on every harness adapter. Images go with text in the standard message format, a content array of text and file parts, and the harness's vision model processes them. The adapters are Claude Code, Cline, Codex, Cursor, Deep Agents, fx, GitHub Copilot, Grok Build, OpenCode, and Pi.
Source for HarnessAgent accepts image inputs (opens in a new tab)GPT-6.1 Terra on AI Gateway
On October 26, 2026, OpenAI's GPT-6.1 Terra became available on Vercel AI Gateway. Vercel describes it as a fast model for visual reasoning, computer use, and writing, and says it improves on the previous Terra model in visual reasoning, factual accuracy, and coding. It accepts text and images and returns text, with a context window of more than 1M tokens, and supports configurable reasoning effort, structured outputs, tool calling, and implicit and explicit prompt caching. Fast mode is available as openai/gpt-6.1-terra-fast. Use openai/gpt-6.1-terra through the AI SDK, the OpenAI-compatible Chat Completions API, the Responses API, or a coding agent connected to AI Gateway. Vercel says OpenAI still recommends its larger GPT-6 models for the most complex coding tasks. AI Gateway reflects provider pricing with no markup.
Source for GPT-6.1 Terra on AI Gateway (opens in a new tab)
OpenCode
OpenCode 2.2.3
Released October 30, 2026. Install Claude Code, run claude auth login, and select Claude Agent in /connect to use that login and its models. OPENCODE_CLAUDE_PATH can override the Claude executable. Desktop streaming and long conversations do less rendering work, and the release fixes side-panel tabs, repeated "Session not found" toasts, and Open file over HTTP. Experimental schedules can run Code Mode JavaScript checks, including statically authorized tool calls, persistent state, and custom wake messages. Breaking changes for code that imports @opencode/ai directly do not change the CLI, Desktop, or HTTP API: speech, transcription, and video move from @opencode/ai/effect to @opencode/ai/experimental, including ai.experimental.speech, ai.experimental.transcription, and ai.experimental.video on the Promise client. TestSpeech, TestTranscription, and TestVideo move to @opencode/ai/experimental/testing. Both clients build tools with tool(...), which replaces AI.tool and Tool.make, and testing.tool is now testing.toolCall. Separately, @opencode/plugin now requires @opentui/core and @opentui/solid 0.8.0 or later. In OpenTUI 0.8, a ScrollBox's wrapper, viewport, content, and scrollbars are facade objects instead of renderables, and a custom RenderContext must implement clock and onAnimationTick. Managed policy integration.use now controls integration accounts only. The new mcp.use, skill.use, and plugin.use actions control MCP servers, skills, and plugins. Rules with the old mcp:, skill:, and plugin: resource prefixes are converted automatically, but a bare integration.use rule no longer applies to those resources.
Source for OpenCode 2.2.3 (opens in a new tab)OpenCode 2.2.4
Released October 30, 2026. Desktop terminals now belong to the session, survive server restarts with their scrollback, and stay synchronized with the TUI and other windows. Renamed terminals keep their names across restarts. With experimental.artifacts enabled, agents can publish, update, and serve HTML or static artifacts through the built-in local provider or plugin-provided storage. Selecting Claude Agent from the provider picker copies claude auth login, can install Claude Code, and checks the connection before model selection. While creating a worktree, Choose base branch is available from the command palette and Settings, Shortcuts, bound to Command-Option-Period on macOS or Ctrl+Alt+Period on Windows and Linux. Enabled schedules run under opencode serve, and multiple server processes coordinate so a schedule fires only once. Schedules stay opt-in through experimental.schedules.
Source for OpenCode 2.2.4 (opens in a new tab)OpenCode 2.2.5
Released October 30, 2026. Command-Option-O on macOS, or Ctrl+Alt+O on Windows and Linux, chooses a worktree, and Command-Option-B or Ctrl+Alt+B chooses its base branch. Command-Shift-X or Ctrl+Shift+X toggles shell mode in the composer. Holding Command or Ctrl in the quote popup, or pressing Command-Enter or Ctrl+Enter, sends that quote and comment and leaves the rest of the composer untouched. compaction.buffer accepts a percentage of the model's context window, such as 20%, in addition to a token count. A tool call named bash runs as shell, with the same permissions, unless a plugin tool is actually named bash. MCP servers that send stdio messages larger than 10 MiB stay connected, and those messages are read much faster.
Source for OpenCode 2.2.5 (opens in a new tab)OpenCode 2.2.1
Released October 29, 2026. Run /goal with an objective in the TUI or Desktop and the session continues until a reviewer confirms the objective is met. /goal pause, /goal resume, and /goal clear control it. Enable experimental.image_generate so agents can create or edit images with the image_generate tool, and optionally set experimental.image_generate.model. Opt-in scheduled prompts are available through /schedule, agent tools, and experimental.schedules. In Desktop, selecting a running shell in the activity pill shows its live output and can stop it, and a redesigned model picker that searches favorites, recent models, providers, and variants is now the default. Codex sign-in can offer GPT Image 2, and MiniMax and Meta Muse image generation are supported. OpenRouter, DigitalOcean, and Venice gain Responses API support, including freeform tools. Breaking changes affect code that imports @opencode/ai directly, not the CLI, Desktop, or HTTP API: the fal, Black Forest Labs, Replicate, and Stability providers and their entry points are removed; five legacy provider entry points are removed in favor of the nested paths 2.2.0 already exports; and each provider's configure() input type is named Config.
Source for OpenCode 2.2.1 (opens in a new tab)OpenCode 2.2.2
Released October 29, 2026. When a connected provider supports image models, agents can use experimental_image_generate without extra configuration. Set a preferred model with experimental.image_generate.model. On the Desktop new-session screen, clicking the current Git branch opens that project's "Run session in" worktree picker. Timeline quotes, terminal quotes, and browser annotations share a compact optional-comment editor, and source highlights stay attached to local drafts. Classic question cards are the default again. Settings, Experimental, New question UI turns on the redesigned keyboard-oriented dock. The SDK adds if and until inputs for client.experimental.schedule.create, plus schedule check results and output fields.
Source for OpenCode 2.2.2 (opens in a new tab)OpenCode 1.19.5
Released October 28, 2026. Failed TUI worker requests now return an error instead of leaving the caller hanging.
Source for OpenCode 1.19.5 (opens in a new tab)OpenCode 2.2.0
Released October 27, 2026. Experimental Desktop plugins can be created, installed, and shared, and managed from Settings, Extensions, Plugins. Plugins can attach and update keyed session instructions with ctx.session.instructions.entry.put, list, and remove; each change is added as a small update at the next step instead of editing the system prompt, and the entries survive compaction and reverts without starting an extra model turn. Experimental image and decision generation APIs are available to plugins and the SDK. Prompt-cache TTLs can be set per provider and model, and the Todo tool is restored in the desktop app as an opt-in. Breaking changes: the @opencode/ai package root exports the Promise API and the Effect API moves to @opencode/ai/effect, @opencode/ai/promise is removed, the OpenCode GitHub Action is no longer published, and TUI plugins require OpenTUI 0.7.0. Sign in with ChatGPT adds GPT-6.1 Terra and GPT-6.1 Terra Fast. The release also refreshes the desktop app.
Source for OpenCode 2.2.0 (opens in a new tab)
Cloudflare Workers
Browser Run uses Chrome 152
On October 30, 2026, Cloudflare Browser Run started running Chrome 152 by default, with no code change required. The default User-Agent for Quick Actions now reports Chrome 152 instead of Chrome 119. A custom userAgent parameter still overrides it. Quick Actions requests look more like a regular Chrome browser. Pages that return large response headers now load correctly. Browser Run removed distrusted root certificates from its trust store. The /links endpoint returns absolute URLs, and links-only crawls follow relative links. /markdown and crawl Markdown output keep code block languages. Quick Actions that time out return a consistent error, and outbound Workers that throw an exception return an HTTP error to the browser.
Source for Browser Run uses Chrome 152 (opens in a new tab)Durable Objects defer code updates by default
On October 30, 2026, Cloudflare Durable Objects began deferring code updates by default. An active Durable Object can keep running its current code for up to 5 minutes (300 seconds) while it waits to hibernate, so in-flight work can finish. WebSocket connections that use the Hibernation API remain connected when an object adopts new code after hibernating. code_update_strategy in the Wrangler durable_objects config can set mode to deferred or immediate and max_delay in seconds, from 0 to 86400. Omitting max_delay uses 300 seconds. A single deployment can pass --durable-objects-code-update-mode immediate or deferred to wrangler versions deploy. A deferred update is best effort and does not guarantee that every operation finishes. The object can still reset before max_delay, for example during a Cloudflare runtime update.
Source for Durable Objects defer code updates by default (opens in a new tab)Agents SDK skills are workspace files
On October 29, 2026, Agent Skills in the Cloudflare Agents SDK became files. A folder with a SKILL.md under /workspace/.agents/skills/ in an agent's Computer Workspace is available on the next turn, with no build or deploy, and the agent can write its own skills. Skills can also load from an R2 bucket or from code, and Think lists every source again before each turn. Every SKILL.md is checked against the Agent Skills specification. A skill that breaks a rule still loads, with a logged warning. Skill scripts are Workspace files, and the model runs them with the Workspace exec tool. The same activate_skill and read_skill_resource tools work in Think, Pi Durable, and the AI SDK. When a skill changes mid-conversation, Think sends the new catalog as a reminder instead of rebuilding the cached system prompt. This is a breaking change for code that uses agents/skills directly. The agents:skills Vite import, agents/skills/compile, SkillRegistry, and Think's skillWorkspace are removed, and runner() moved to agents/skills/runner. Install Agents SDK v0.32.0 and Think v0.23.0 or later.
Source for Agents SDK skills are workspace files (opens in a new tab)Agents SDK browser previews without public URLs
On October 28, 2026, Cloudflare Agents SDK 0.31.0 or later let an agent's persistent browser open internal hostnames such as preview.internal. Requests to those hostnames go to the agent instead of the public Internet, so the model can check a dev server in its sandbox, a folder in its workspace, or another Worker without exposing that server publicly. Each host's fetch function runs in the Durable Object and can use its bindings, or the host can send requests to another Worker through a service binding. A target schema lets the model choose what the host serves, such as a port, through browser.assignHost(). The checked target lasts across turns, browser restarts, and redeploys. Built-in browser-tool adapters cover Pi Durable, the AI SDK, and TanStack AI. Another harness can implement a tool with browser.session(). browserTool now runs the model's code in its own Dynamic Worker and no longer needs a CodemodeRuntime export.
Source for Agents SDK browser previews without public URLs (opens in a new tab)Cloudflare Computer can resume and wait for approval
On October 26, 2026, @cloudflare/computer 0.7.0 can resume agent code after a deploy or crash and pause for an external action such as approving a tool call. Set onInterrupt to "replay" on WorkerJavaScriptBackend so interrupted JavaScript runs again from the beginning and reuses recorded results for completed calls. A call whose result was not recorded before the interruption may run again. A custom module function can call ctx.interrupt() to pause until the app supplies a decision. Time spent waiting does not count toward the timeout, and the wait survives deploys. Pending calls are listed with workspace.runtime.pending() and resolved with respond().
Source for Cloudflare Computer can resume and wait for approval (opens in a new tab)
Claude Code
Claude Code 2.1.316
Released October 31, 2026. CLAUDE_CODE_DISABLE_ULTRATHINK turns the ultrathink keyword off. The Claude apps gateway adds pooled group budgets: one monthly budget that a group's members share, on top of each member's own limit, set with pooled_budget on the spend limits admin API. CLAUDE_CODE_SUBAGENT_EFFORT sets the effort level subagents run at when nothing else names one, instead of inheriting the session's. managed.policy_mode set to layered applies every matching gateway policy, each merged over the ones listed below it. guardrail.input_tagging on Bedrock upstreams wraps prompts in the input tags an Amazon Bedrock guardrail's prompt attacks filter needs, and all_messages_and_system also tags the system prompt. Fast mode for Opus 4.8 is deprecated as of October 30, 2026, and will be removed on November 30, 2026. After removal, fast mode requests on Opus 4.8 return an error and do not fall back to standard Opus 4.8. Migrate to Opus 5.5 to keep the speedup. VS Code can sign in to a Claude apps gateway from the chat panel, and ultrathink runs its turn at xhigh effort. Cloud sessions can fill in or decline a form an MCP server asks for, and Enterprise members can add and delete network secrets on their own cloud environments. The Code Review analytics page adds a findings-by-severity card. The release also includes a large fix sweep for the shell, permissions, MCP, cloud sessions, VS Code, Claude Tag, and the terminal UI.
Source for Claude Code 2.1.316 (opens in a new tab)Claude Code 2.1.314
Released October 30, 2026. permissions.blockWritesOutsideWorkingDirectories makes Write, Edit, and NotebookEdit refuse paths outside the working directories in every permission mode. Claude apps gateway upstreams can list IdP groups so only those developers are sent to that upstream, including on failover, and Bedrock upstreams can set guardrail_trace to log which Amazon Bedrock guardrail filter fired without the matched text. In VS Code and Claude Desktop's Code tab, a conversation of 200K or more tokens that is left idle is summarized, and that summary is used on the next message if the prompt cache has expired. Idle compaction also runs for subscribers with telemetry turned off, and on Bedrock, Vertex, Foundry, and LLM gateways once the one-hour prompt cache is on. File tools ask before editing the local copy of an organization's server-managed settings. Flagged requests retry on another model instead of ending in a refusal when ANTHROPIC_DEFAULT_OPUS_MODEL, availableModels, or deniedModels leaves only an older Opus model, or none. Cloud sessions from the web, desktop, and mobile apps start even when a setup script exits with an error. The release also includes a large fix sweep for permissions, MCP sign-in, Remote Control, cloud sessions, VS Code, Claude Tag, and the terminal UI.
Source for Claude Code 2.1.314 (opens in a new tab)Claude Code 2.1.312
Released October 27, 2026. The Claude apps gateway adds oidc.user_id_claim so a person can be identified by an id_token claim other than sub, such as Microsoft Entra ID's oid. In fullscreen, clicking the "to send now" hint under a queued message sends it right away. Write now asks Claude to read all of a very large CLAUDE.md, rules file, or @-included file in a subdirectory, or to edit it in place, instead of overwriting it after seeing only the start. Bedrock, Vertex, Foundry, and Mantle use the shorter system prompt for model names Claude Code does not recognize, such as custom deployment names and inference profile ARNs; CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0 keeps the full prompt. Cloud routines can set an effort level, and the cloud environment setup script box has a Bash editor. Claude Code stops sending the claude_code.safety_event OpenTelemetry event. The release also includes a large fix sweep for skills, commands, VS Code, cloud sessions, Remote Control, Claude Tag, and Code Review.
Source for Claude Code 2.1.312 (opens in a new tab)Claude Code 2.1.311
Released October 26, 2026. self-hosted-runner adds --hold-idle-file and --max-hold-min to keep an idle session while the host reports processes it left running. In the interactive terminal, ultrathink runs its turn at xhigh effort on Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.1 through Anthropic's API, and typing it while Claude is working raises the rest of that turn. A standard-context session whose request runs over 200K tokens compacts and continues instead of stopping with "Usage credits required for 1M context". Claude Code on Bedrock, Mantle, Vertex, and Foundry loads model settings from a signed file on downloads.claude.ai; CLAUDE_CODE_MODEL_CATALOG=off turns that off. Claude Tag analytics adds a Skills view for how often each attached skill runs in Slack channels. The release also includes a large fix sweep for startup freezes, MCP, compaction, subagents, VS Code, cloud sessions, and Code Review.
Source for Claude Code 2.1.311 (opens in a new tab)
Figma
Google Drive connector for the Figma agent
On October 30, 2026, Google Drive could be connected to the Figma agent and Make, on every Figma plan including Starter. The connector brings existing file context into Figma and can create new Drive files without leaving the file. Figma says it can summarize research insights, reference data, or pull in assets to edit on the canvas. Connectors are added from the plus menu in the prompt box when collaborating with Figma agents.
Source for Google Drive connector for the Figma agent (opens in a new tab)Internal publishing for Weave tools
On October 30, 2026, Figma Organization and Enterprise plans could publish Weave tools for use within the organization. Internally published Weave tools appear in the Tools tab alongside Weave tools built by Figma and the Figma Community.
Source for Internal publishing for Weave tools (opens in a new tab)Version history for generative plugins and shaders
On October 27, 2026, editing a generative plugin or shader in Figma started saving a new version, so earlier versions can be revisited and restored. Each version gets an agent-generated description that forms a change log and shows which version is published. Figma says this is rolling out this week.
Source for Version history for generative plugins and shaders (opens in a new tab)Figma Weave video and image models in the Figma MCP
On October 26, 2026, the video and image models available through Figma Weave became accessible from an external AI agent via the Figma MCP. With a Figma Weave account connected to a Figma account, the agent can generate videos and images with models such as Seedance, Kling, and Nano Banana while chatting. Videos and images generated with Weave models consume Weave credits.
Source for Figma Weave video and image models in the Figma MCP (opens in a new tab)
Devin Desktop
Devin Desktop 3.11.235
Released October 30, 2026. Downloaded diagnostics now keep the last 4 MB of each output log, up from 25 KB, so an export includes the failure being reported.
Source for Devin Desktop 3.11.235 (opens in a new tab)Devin Desktop 3.11.238
Released October 30, 2026. Members an enterprise admin has listed for Anthropic's Cyber Verification Program can switch the program on or off for themselves from the Devin Local model picker. Off routes their Anthropic requests like everyone else's, under the team's usual data retention settings.
Source for Devin Desktop 3.11.238 (opens in a new tab)Devin Desktop 3.11.228
Released October 26, 2026. In Remote-SSH and Remote-WSL windows, the Agent Window folder picker sets the session's working directory from the window's remote folders and no longer offers local ones, and file links in the Agent Window open from the remote machine. The status bar no longer disappears after switching from Agent mode to IDE mode. On macOS, banners at the top of the window no longer interfere with the window controls. IDE mode offers Devin Cloud whenever the account has Cloud access, the same as the Agent Window, and the "Devin for Cloud" team setting is removed. The "Connect your GitHub account" callout has a "Don't notify me again" option. A dropped connection during sign-in no longer signs you out; Devin Desktop retries instead. Devin Local sessions load again after an update that previously left them failing with "Failed to list sessions".
Source for Devin Desktop 3.11.228 (opens in a new tab)
Gemini API
Gemini Computer Use generally available
On October 30, 2026, Google made the Computer Use tool generally available across the Interactions API and the GenerateContent API.
Source for Gemini Computer Use generally available (opens in a new tab)Gemini API CLI beta
On October 27, 2026, Google released the Gemini API CLI (gemini-api) in beta. It is a stateless, zero-dependency command-line client for AI coding agents, scripts, and CI pipelines. It uses the Interactions API to generate text and media, analyze files, and run Gemini managed agents.
Source for Gemini API CLI beta (opens in a new tab)Generate Content API stores requests by default
On October 26, 2026, new projects using the Gemini Generate Content API (generateContent, streamGenerateContent, and batchGenerateContent) on the paid tier store requests by default (store=true), matching the Interactions API. Existing projects keep their current setting. To opt out, set store=false on a request or turn off Generate Content API logging in the Settings panel of the Logs and Datasets dashboard.
Source for Generate Content API stores requests by default (opens in a new tab)
Anthropic API
Opus 4.8 fast mode deprecated
On October 30, 2026, Anthropic deprecated fast mode for Claude Opus 4.8, with removal on November 30, 2026. After removal, requests to claude-opus-4-8 with speed set to fast return an error. To continue using fast mode, migrate to Claude Opus 5.5.
Source for Opus 4.8 fast mode deprecated (opens in a new tab)Usage and Cost Analytics API
On October 29, 2026, Anthropic launched the Usage and Cost Analytics API in beta with the analytics-usage-cost-2026-08-24 beta header. Its /v1/organizations/analytics/usage and /v1/organizations/analytics/cost endpoints report token usage and cost for Claude Console and Claude Enterprise organizations. Results can be filtered and grouped by dimensions such as model, workspace, API key, and the user or service account behind each request. The Usage and Cost API is now a legacy API, and so are the usage_report, cost_report, user_usage_report, and user_cost_report endpoints of the Claude Enterprise Analytics API. All of them remain supported.
Source for Usage and Cost Analytics API (opens in a new tab)
Codex CLI
Codex CLI 0.166.0
Released October 29, 2026. The TUI can select transcript text from the keyboard, with word, line, and page motions, integrated search, and configurable shortcuts, and it renders Markdown consistently in async questions, choices, answers, and transcript replies. GPT-6.1 Terra is added to the bundled, Vertex AI, and Amazon Bedrock model catalogs. codex queue --resume resumes unloaded threads after a message is queued. Safety blocks in the TUI explain the block more clearly and offer selectable safer follow-up prompts. codex doctor --network-only checks configuration, connectivity, and WebSocket diagnostics. The release also preserves acknowledged approval and permission settings when resuming or forking threads, keeps local image attachments when restoring killed composer text, fixes standalone release cleanup on macOS when installations are in use or paths contain non-ASCII characters, restores spawned descendant threads when their parent is unarchived, fixes app-server thread startup and shutdown races, improves TUI reconnection including offline /quit and /exit, and preserves reasoning-token accounting across resume and fork so compaction is not premature. The standalone executor ships in the canonical CLI and app-server release packages.
Source for Codex CLI 0.166.0 (opens in a new tab)Codex CLI 0.165.0
Released October 27, 2026. Adds a built-in Google Cloud Vertex AI provider, configured with model_provider = "gcp", and adds GPT-6.1 Terra to the model catalogs. The session picker can filter by repository, and the agent command center groups sessions by directory across linked worktrees. Reasoning effort can change during a conversation by default for supported models. Memory v2 is on by default for fresh profiles, and external-agent memories can be imported into it. The app-server API adds experimental attached thread forks. The release also fixes TUI approval defaults, MCP cancellation, queued messages after an interrupt or a switch to Plan mode, model-not-found retries, and a marketplace URL check that could let Git connect to a different host than the one checked against allowed patterns.
Source for Codex CLI 0.165.0 (opens in a new tab)
Gemini CLI
Gemini CLI 0.66.0
Released October 27, 2026. Terminal output is synchronized, the spinner renders on its own, and scrollback is kept across viewport refreshes. The release stops buffer flickering and border bleed in scrollable tool containers. The scheduler emits tool-call groups directly, drops UI lookahead, and forwards termination signals from bootstrap wrappers to child processes. Shell and git environments keep the user-level ~/.gitconfig, and interactive SSH and credential prompts are suppressed. MCP OAuth can opt in to private-network endpoints. Settings placeholders resolve after .env files load, and keybindings.json accepts a keys array. A banner says Gemini CLI is not optimized for Gemini 4.
Source for Gemini CLI 0.66.0 (opens in a new tab)
Mistral
Mistral Large 4 open weights
On October 29, 2026, Mistral released the weights of an updated Mistral Large 4, also called Le Chonk, after the earlier public preview. Mistral says further reinforcement learning improved coding and defensive cybersecurity, and that reasoning efficiency lowered inference cost. The model adds three reasoning modes: Low for simple tasks, which replaces None, Medium for everyday tasks, and High, the default, for the hardest cases. It is available in Mistral Studio, through the Mistral API, and on OpenRouter. The weights can be downloaded from Hugging Face and served with vLLM or SGLang. The Mistral Ecosystem Open-Weight license, a modified MIT license, allows use, modification, fine-tuning, and commercialization, and a business that operates a model-as-a-service offering needs a separate commercial license from Mistral. API pricing is $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. Mistral says the model scores 19.7% on Cognition's FrontierCode 2.0 and 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task on average. It says Terminal-Bench 4.0 rose 13 percentage points from the public preview to 41%, with 30% fewer output tokens, and that the Artificial Analysis Cyber Index score is 51%, with a 96.6% average refusal rate on the malicious cyber prompts it measures.
Source for Mistral Large 4 open weights (opens in a new tab)
Netlify
GPT-6.1 Terra in AI Gateway and Agent Runners
On October 26, 2026, OpenAI's GPT-6.1 Terra became available through Netlify's AI Gateway and Agent Runners. Netlify Functions can call it with the OpenAI SDK without a separate API key; the changelog example uses the Responses API with model gpt-6.1-terra. It is also available in Scheduled Functions, Background Functions, and Edge Functions, and those calls use Netlify's caching, rate limiting, and authentication.
Source for GPT-6.1 Terra in AI Gateway and Agent Runners (opens in a new tab)
OpenAI API
GPT-6.1 Terra
On October 26, 2026, OpenAI released GPT-6.1 Terra. It is available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and in the API as gpt-6.1-terra. OpenAI describes it as a fast model for visual reasoning, computer use, and writing. Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens, which OpenAI says is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. For the most complex coding tasks, OpenAI says GPT-6.1 Sol and GPT-6 Astra remain the better choice.
Source for GPT-6.1 Terra (opens in a new tab)
Replit
Cloud Agents for existing repositories
On October 27, 2026, Replit made Cloud Agents for existing repositories available to all Pro and Enterprise customers. Someone describes a change in plain language. The agent finds the relevant code, updates a live preview, and opens a GitHub pull request or a GitLab merge request that follows the team's skills and contribution guidelines. It can test the change in an isolated cloud sandbox by clicking through the running app, then attach screenshots and video to the request. Review feedback worth keeping can become a proposed skill update that an engineer approves before it takes effect; that part is in early access. An admin can share one environment: a Cloud Agent configures a virtual machine from the codebase, and it updates from the main branch. Mention or DM @Replit in Slack to start from a thread. The Slack integration is available to all customers, and a pull request can also be started from the Replit mobile app. Microsoft Teams support is coming soon. Connect a repository at replit.com/repos.
Source for Cloud Agents for existing repositories (opens in a new tab)
Zed
Zed 1.26.2
Released October 28, 2026. Adds Pyrefly as an optional built-in Python language server, expandable debugger hovers for nested variable values, and DWARF debugging for WebAssembly sources in the JavaScript debugger. ACP configuration defaults now apply only to new conversations. Agent model and configuration pickers search more accurately for Zed Agent and third-party ACP agents. OpenCode replay handles DeepSeek reasoning and can disable thinking for Google models, and thinking controls explain when thinking cannot be disabled at the selected effort level. The release also fixes hangs when restoring agent threads with subagents, a crash in the Agent Panel when a mentioned file name contains a line break, agent authentication overwriting drafts, duplicated thinking text from OpenAI-compatible providers that send both reasoning and reasoning_content, OpenCode requests failing for Claude models that use adaptive thinking block binding, and screenshots pasting twice into the Agent Panel or inline assistant.
Source for Zed 1.26.2 (opens in a new tab)
Updated