Skip to content
StackPick

What shipped · 2026-w47

What shipped in AI dev tools: week of Nov 16, 2026

Nov 16, 2026 – Nov 22, 2026

88 releases across 23 tools this week; biggest: GitHub Copilot (26 releases).

  1. Copilot connectors for Supabase, Box, Calendly, and more

    On November 18, 2026, GitHub Copilot added eight GitHub-verified MCP servers, or connectors, in public preview: Box, Buildkite, Calendly, ClickUp, Mixpanel, Supabase, Webflow, and Wix. They join more than 25 services already available, such as Slack, Figma, and Sentry. Connect a service once and use it across Copilot. In the Copilot app, open Customize and select Connect. In Copilot CLI, run /mcp, open the Connectors tab, and select Connect. In VS Code, open the Agents window, go to Customizations, and connect the service. On github.com, browse services on the Copilot customization page. Connectors are in public preview for all Copilot plans. With Copilot Business or Copilot Enterprise, an admin needs to enable the Editor preview features policy. If the enterprise MCP allowlist lists connectors individually rather than with a wildcard, an admin also needs to add the new connectors. Some services also need an admin to enable them on that service's side, such as allowing Box MCP access.

    Source for Copilot connectors for Supabase, Box, Calendly, and more (opens in a new tab)
  2. Gemini 4 Neon in GitHub Copilot

    On November 18, 2026, Gemini 4 Neon became generally available in GitHub Copilot. GitHub says it is designed for agentic coding, multistep tool use, and knowledge work. According to Google, Gemini 4 Neon delivers quality close to Gemini 4 Carbon at a lower token cost per task, starts responding faster, makes fewer unrequested changes on coding tasks, and is less likely to make unsupported claims about code, tool, or test results. It is billed at provider list pricing under usage-based billing. It is available to Copilot Pro+, Max, Business, and Enterprise users in the model picker in Visual Studio Code, Visual Studio, Copilot CLI, the GitHub Copilot coding agent, the GitHub Copilot app, github.com, GitHub Mobile on iOS and Android, JetBrains IDEs, Xcode, and Eclipse. Rollout is gradual. Copilot Business and Copilot Enterprise administrators manage access through the model policy in Copilot settings. Under default model enablement, new models are automatically enabled unless an administrator has turned off the global default or explicitly disables this model.

    Source for Gemini 4 Neon in GitHub Copilot (opens in a new tab)
  3. GitHub CLI waits for pull request events

    On November 18, 2026, GitHub CLI 2.105.0 could wait on GitHub Event Streams. gh event listen takes a pull request URL and the events to wait for, such as a submitted review, creates the subscription, opens the stream, and prints each event as a line of JSON. By default it exits after the first event. --count waits for several events, --follow keeps listening, and --timeout stops waiting. --jq or --template selects fields. It can wait for submitted reviews, pull request state and head changes, merge conflicts, and completed check suites. gh event subscription create, list, view, and delete manage subscriptions directly. gh pr create, gh pr view, gh run view, gh run watch, and gh api suggest gh event listen instead of polling. The gh event commands are in preview and may change. They require Event Streams, which is in preview on GitHub.com once the waitlist approves access.

    Source for GitHub CLI waits for pull request events (opens in a new tab)
  4. GitHub Copilot CLI 1.0.114

    Released November 18, 2026. The published notes say fixes and changes. The only file changed from 1.0.113 is changelog.md, updated with the 1.0.113 notes.

    Source for GitHub Copilot CLI 1.0.114 (opens in a new tab)
  5. Team defaults for Copilot plugins

    On November 18, 2026, team customizations for GitHub Copilot plugins became generally available for organization teams on Copilot Business and Copilot Enterprise. Team maintainers and organization owners can select approved Agent Plugins as team defaults, between individual customization and organization-wide policy. On the organization team page, the Customize Copilot tab lists plugins from the marketplaces built into Copilot, the enterprise's marketplaces, and any marketplace the team adds from a GitHub repository. Selected plugins are enabled by default for the team's direct members in VS Code, the GitHub Copilot app, and Copilot CLI. Members of child teams do not inherit them. Team defaults are optional, and members can turn a team plugin off where policy permits. Enterprise and organization policies take precedence. Copilot clients still enforce which marketplaces are allowed and which plugins are required or blocked.

    Source for Team defaults for Copilot plugins (opens in a new tab)
  6. Adaptive Copilot code review and secret detection

    On November 17, 2026, GitHub Copilot code review added Adaptive effort, cross-repository context, and secret detection, in public preview, for Copilot Pro, Pro+, Max, Business, and Enterprise. Adaptive looks at a pull request's complexity and risk, reviews it with Lite, Balanced, or Max, and shows which level it picked. Select it for one pull request from the Reviewers bar, or set it as the default in personal, repository, organization, or enterprise settings. Balanced remains the default. Adaptive never picks a level above the maximum configured for the enterprise, organization, or repository. Cross-repository context, for public repositories, lets a review read related public repositories, including code that depends on the change, documentation, custom instructions, and skills. Secret detection flags exposed secrets in review comments, using deterministic detectors for more than 500 known secret types from GitHub's partner program and an LLM plus GitHub's purpose-built model for secrets without a recognizable format. Turn it on under Copilot, Code review, in repository or organization settings. The same note says repository insights can show how review feedback was addressed, and that Copilot code review for Azure Repos is generally available. Azure administrators can enable it at the organization, project, and repository level, set Lite or Balanced as a default, use custom instructions and skills, apply a suggestion and resolve its comment together, and see audit events when settings change. Usage is billed in GitHub AI credits to the Azure subscription linked to the Azure DevOps organization.

    Source for Adaptive Copilot code review and secret detection (opens in a new tab)
  7. Copilot activity in Microsoft Agent 365

    On November 17, 2026, enterprises could stream GitHub Copilot agent activity into Microsoft Agent 365. Administrators can see activity across Copilot clients, including cloud agents on GitHub.com, Copilot CLI, Visual Studio Code, Visual Studio, and partner IDEs such as JetBrains and Eclipse. They can review user sessions, including prompts, responses, and tool calls, and view usage trends in Agent 365. Additional risk information is available in Microsoft Purview and Defender. This is available for GitHub Enterprise Cloud customers with Enterprise Managed Users.

    Source for Copilot activity in Microsoft Agent 365 (opens in a new tab)
  8. Copilot app Azure DevOps and browser use

    On November 17, 2026, the GitHub Copilot app could work with Azure DevOps repositories in public preview. Clone a repository, ask an agent to create a branch, change code, and open a pull request, then track it in the sidebar. Fix on a review comment sends an agent to address it, and agent merge can fix CI failures and complete the pull request once Azure DevOps checks and policies pass. Turn it on under Settings, Experimental, then add the Microsoft account under Accounts. Cloud sessions are not available for Azure DevOps repositories. Browser use, in public preview, works in Chromium-based browsers including Chrome and Edge with the GitHub Copilot extension for Chrome, in an isolated tab group, and can be used with computer use. It is off by default. Install the extension, then turn on Enable Computer Use and Enable Browser Use under Settings, Computer Use. Browser use is also available in Copilot CLI, VS Code, and JetBrains IDEs. The same note says GitHub Enterprise Server support in the app, previously an opt-in preview, is generally available in GitHub Copilot app 1.2.0 or later, and that cloud sessions are not available for GitHub Enterprise Server.

    Source for Copilot app Azure DevOps and browser use (opens in a new tab)
  9. Copilot benchmarks compare harnesses

    On November 17, 2026, the GitHub Copilot benchmarks site added Terminal-Bench-Science 0.1 and a Harnesses view, both in public preview. Terminal-Bench-Science 0.1 evaluates agents on research workflows across the life, physical, earth, mathematical, and engineering sciences, with tasks authored and reviewed by domain experts. The Harnesses view compares GitHub Copilot CLI with Claude Code and Codex CLI using the same model at the same reasoning effort. A scatter plot places each harness by resolution rate against mean cost per task, and paired differences show Copilot CLI's resolution advantage and cost savings for each matched model, with approximate 95% bootstrap intervals. Only exact model pairs at medium reasoning effort are shown. Copilot-only configurations, including GitHub Copilot Fusion, and unmatched results are excluded.

    Source for Copilot benchmarks compare harnesses (opens in a new tab)
  10. Copilot browser use with the Chrome extension

    On November 17, 2026, browser use became available in public preview in GitHub Copilot CLI, the GitHub Copilot app, VS Code, and JetBrains, paired with the GitHub Copilot Chrome extension. It can be used with computer use. Install GitHub Copilot in Chrome, then enable browser use: in the CLI, open /plugin and enable browser use, then start a new session. In the app, turn on Enable Browser Use under Settings, Computer Use. In VS Code and JetBrains, turn on Enable Browser Use in Settings. The extension should show Connected. Its settings manage permission levels and experimental features.

    Source for Copilot browser use with the Chrome extension (opens in a new tab)
  11. Copilot CLI Microsoft Entra sign-in preview

    On November 17, 2026, GitHub Copilot CLI added Microsoft Entra authentication in public preview. It requires Copilot CLI 1.0.113 or later. On Windows and macOS, you can sign in with a Microsoft Entra account to authenticate silently to a linked GitHub Enterprise Managed User account, use a Microsoft 365 Copilot license, or both, depending on the organization's entitlements. Linux support will follow. For enterprises using Enterprise Managed Users via OIDC, silent SSO uses the Entra identity without a separate GitHub sign-in. That path requires an assigned Copilot Business or Copilot Enterprise license and the Copilot CLI policy, uses GitHub inference, and is limited to Copilot CLI and VS Code. If the organization is in Frontier and has a Microsoft 365 Copilot license with usage-based billing, Entra sign-in can use models through Microsoft Copilot from the terminal without a GitHub account. That path uses Microsoft 365 inference, and model access is managed in Microsoft 365 Copilot settings. The Work IQ MCP server can add Microsoft 365 work data as context.

    Source for Copilot CLI Microsoft Entra sign-in preview (opens in a new tab)
  12. Copilot code review outcomes in Repository insights

    On November 17, 2026, a repository's Insights tab added Repository insights as its first item, for Copilot code review outcomes. Anyone with read access can see comments addressed, high-severity comments addressed, pull requests with no later human change request, and merged pull requests without a Copilot review. The page also shows the addressed rate, with numerator, denominator, unevaluable exclusions, trend, and breakdowns, and the categories of feedback Copilot leaves. The same metrics are in the enterprise and organization repository 1-day reports, repos-1-day, in the Copilot usage metrics API. Enterprise owners and billing managers, organization owners, and anyone with a custom role that grants View Copilot Metrics can use those API reports. The Copilot usage metrics policy must be enabled.

    Source for Copilot code review outcomes in Repository insights (opens in a new tab)
  13. Copilot Fusion on more surfaces, with new models

    On November 17, 2026, GitHub expanded Copilot Fusion. GitHub says Copilot Fusion Max reached a 71.2% resolution rate on the public Terminal-Bench 4.1 leaderboards hosted by Stanford, Harbor, and the Laude Institute, 5.4 percentage points ahead of Opus 5.5 and Fable 5.2, while costing 88% less than Fable 5.2 and 57% less than Opus 5.5. Copilot Fusion can be selected on the latest github.com, Visual Studio, and GitHub Mobile beta apps for iOS and Android. Support in the production GitHub Mobile apps is coming shortly. This release adds Claude Sonnet 5.5 to Copilot Fusion Max, GPT-6.1 Terra to Copilot Fusion High, and MAI-Code-2 to Copilot Fusion Standard.

    Source for Copilot Fusion on more surfaces, with new models (opens in a new tab)
  14. Copilot in JetBrains, Ignite releases

    On November 17, 2026, GitHub Copilot for JetBrains added cloud sandbox, remote access, browser use, and on-device voice dictation, each in public preview. Cloud sandbox runs sessions in an isolated GitHub-hosted environment, rolling out in phases. Remote access can make chats available in other Copilot clients such as VS Code and the Copilot app, also rolling out in phases. Browser use, paired with GitHub Copilot in Chrome, opens and navigates pages and interacts with links, buttons, and forms. Enable it under GitHub Copilot, Chat, Enable Browser Use. On-device voice dictation is available on supported native plugin packages with Editor Preview enabled. Transcripts stay as editable drafts until you select Send. Split mode remote development runs chat and editor UI on the local client and Copilot operations on the remote backend, which requires the plugin in both places. The update adds timestamps, session delete beside archive, Copilot Fusion pinned at the top of model management, organization usage in the quota dialog, Enter to steer a running session and Ctrl+Enter or Command+Enter to queue a follow-up, shorter worktree names, and inline model and token details. Code review uses the Copilot agent review engine in the existing inline interface. Java modernization findings appear in the Problems view on supported IDEs, and opening or importing a Maven project checks dependency coordinates with GitHub Advisories. Enterprise-managed settings disable Claude and Codex unless an administrator policy allows an exception, and Copilot Fusion applies enterprise-managed tier defaults while keeping available tiers selectable. Support for JetBrains IDE 2025.2 has ended. Upgrade to 2025.3 or later. GitHub recommends 2026.2 or later.

    Source for Copilot in JetBrains, Ignite releases (opens in a new tab)
  15. GitHub Copilot CLI 1.0.113

    Released November 17, 2026. Administrators can disable built-in Browser Use with the managed setting features.browserUse, including in sessions that are already running.

    Source for GitHub Copilot CLI 1.0.113 (opens in a new tab)
  16. GitHub Copilot Evals public preview

    On November 17, 2026, GitHub Copilot Evals entered public preview for Copilot Individual, Copilot Business, and Copilot Enterprise. An eval gives an agent a task, describes a good result, and checks the result and how the agent got there, so a skill, custom agent, or MCP tool can be rerun after a change. Ask Copilot or use /eval to create an eval, run it after you approve the run, see which checks passed, and compare runs from before and after a customization change. The evaluation engine is Vally. Install the GitHub Copilot Evals extension from the Visual Studio Marketplace in VS Code, install the plugin in Copilot CLI, or in the GitHub Copilot app open Customize, add the github/evals marketplace, and install the plugin.

    Source for GitHub Copilot Evals public preview (opens in a new tab)
  17. GitHub Event Streams preview

    On November 17, 2026, GitHub Event Streams entered preview. An agent can subscribe to a pull request, keep a connection open, and receive events when a review is submitted, new commits are pushed, a merge conflict appears, the pull request is merged or closed, or a check suite changes, including one event per check suite status change. No public webhook endpoint is required, so an agent on a laptop can receive the events. GitHub says this avoids polling that either burns rate limit or delays review feedback. Event Streams is available on all GitHub plans once the waitlist approves access.

    Source for GitHub Event Streams preview (opens in a new tab)
  18. GPT-6.2 Astra and Sol in GitHub Copilot

    On November 17, 2026, OpenAI's GPT-6.2 Astra and GPT-6.2 Sol became generally available in GitHub Copilot. GitHub says both are for agentic coding and code review, and that in early testing both caught significantly more real issues than earlier GPT-6 models while using fewer tokens. GitHub describes Astra as the latest Astra model for agentic coding and terminal work, and says early testing caught significantly more code-review issues than GPT-6 Astra while completing agentic coding tasks with noticeably fewer tokens. GitHub describes Sol as the latest Sol model for complex agentic coding and in-depth code review, and says early testing caught nearly twice as many real issues as GPT-6.1 Sol while keeping findings accurate. They are available to Copilot Pro+, Max, Business, and Enterprise users in the model picker in Visual Studio Code, Visual Studio, Copilot CLI, the GitHub Copilot coding agent, the GitHub Copilot app, github.com, GitHub Mobile on iOS and Android, JetBrains IDEs, Xcode, and Eclipse. Rollout is gradual. They are billed at provider list pricing under usage-based billing. Copilot Business and Copilot Enterprise administrators manage access through the model policy in Copilot settings. Under default model enablement, new models are automatically enabled unless an administrator has turned off the global default or explicitly disables these models.

    Source for GPT-6.2 Astra and Sol in GitHub Copilot (opens in a new tab)
  19. MAI Code 2 and MAI Code 2 Flash in GitHub Copilot

    On November 17, 2026, MAI Code 2 and MAI Code 2 Flash became available in GitHub Copilot. GitHub describes MAI Code 2 as a general-purpose coding model for deep debugging, multi-file implementation, refactoring, and repository-wide reasoning, and MAI Code 2 Flash as a lighter model for explanations, targeted edits, routine fixes, and rapid iteration. MAI Code 2 is available to Copilot Pro, Pro+, Max, Business, and Enterprise. MAI Code 2 Flash is available to Copilot Free, Student, Pro, Pro+, Max, Business, and Enterprise. Both are available to eligible subscribers on annual plans, 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, Xcode, and Eclipse. They are billed under usage-based billing. Copilot Business and Copilot Enterprise administrators manage access through the model policy. Under default model enablement, new models are enabled unless an administrator has turned off the global default or explicitly disables the model.

    Source for MAI Code 2 and MAI Code 2 Flash in GitHub Copilot (opens in a new tab)
  20. MAI-Code-1.1-Flash retires on December 18

    On November 17, 2026, GitHub announced that MAI-Code-1.1-Flash will be deprecated on December 18, 2026, across all GitHub Copilot experiences, with the launch of MAI Code 2 Flash. The suggested alternative is the latest MAI Code Flash model. Copilot Enterprise administrators may need to enable that model through model policies in Copilot settings. No action is required to remove the older model after it is deprecated.

    Source for MAI-Code-1.1-Flash retires on December 18 (opens in a new tab)
  21. Secret scanning in Copilot security review

    On November 17, 2026, Copilot's security review specialist could run GitHub secret scanning, in public preview, so developers and agents can catch secrets in code changes before they are merged. The review checks the changes, not repository history. Once you opt in, it runs GitHub's detectors for more than 500 secret types from the partner program, and GitHub's fine-tuned model for credentials that do not have a recognizable format, alongside the existing LLM-based review. It is available in Copilot CLI and Copilot app sessions, through Copilot SDK integrations, and in Copilot code review, wherever Copilot security review or Copilot code review is already available. Secret scanning is off by default. In the CLI, turn it on with /settings aiSecretDetection on. In the app, turn on AI secret detection under Settings, Experimental. In the SDK, set the SECURITY_REVIEW_FINE_TUNED_MODEL feature flag to true when creating a session. In code review, turn on Secret scanning under the repository's Settings, Copilot, Code review. Run it by delegating to the security-review specialist, with /security-review, or by requesting a Copilot review on a pull request. Scans consume AI credits under the Secret Protection AI Credit SKU.

    Source for Secret scanning in Copilot security review (opens in a new tab)
  22. Bulk start AI Scan for repositories

    On November 16, 2026, AI Scan for repositories can be started for up to 20 repositories at once, in public preview. Scans can be started for selected repositories when the organization-level setting is enabled, from the code scanning organization alerts page, or for one repository from its alerts page. AI Scan finds security vulnerabilities on a repository's default branch, including languages CodeQL does not support and additional issues in languages CodeQL already analyzes. Findings appear in the code scanning backlog alongside CodeQL alerts. It is available on github.com and GitHub Enterprise Cloud with data residency. A GitHub Copilot license is not required.

    Source for Bulk start AI Scan for repositories (opens in a new tab)
  23. Copilot app phone handoff and minute schedules

    On November 16, 2026, GitHub's Copilot weekly note for the week of November 9 described changes in the GitHub Copilot app. Cloud and Codespace sessions can continue on a phone: Open in includes GitHub Mobile and shows a QR code. When an MCP server such as Slack needs you to sign in again, the conversation offers the matching connector, connects it in place, and resumes the task. Project automations can be scheduled down to the minute, or with a custom cron schedule such as weekdays at 9:30. Dictation can use French, Spanish, and other languages after the multilingual voice model is downloaded. Issues and pull requests can be filtered by custom fields, such as a GitHub Project's Size. High contrast can be combined with red, green, or blue-yellow color filters, and GitHub Dimmed is available as a dark-only theme. The same note also points to Claude Fable 5.2, Gemini 4 Carbon, GitHub Enterprise Server in the Copilot app, and VS Code 1.146, which have their own notes.

    Source for Copilot app phone handoff and minute schedules (opens in a new tab)
  24. Copilot retired three models on November 16

    On November 16, 2026, GitHub deprecated Claude Opus 4.8, Claude Opus 5, and GPT-5.6 Terra across most GitHub Copilot experiences, including chat, inline edits, ask and agent modes, and code completions. These deprecations were announced on October 15, 2026. Suggested alternatives are Claude Opus 5.5, and GPT-6.1 Sol or GPT-6 Luna. Claude Opus 4.8 remains available to individual GitHub Copilot subscribers on annual plans. Claude Opus 5 is removed from the model picker but remains available for runtime refusal fallback through existing authorized access. Existing enterprise model policies continue to apply. Under default model enablement, the suggested alternatives are automatically enabled for Copilot Business and Copilot Enterprise unless an administrator has turned off the global default or explicitly disabled the model. No action is required to remove the deprecated models.

    Source for Copilot retired three models on November 16 (opens in a new tab)
  25. Plugin-only Copilot customizations

    On November 16, 2026, GitHub Enterprise administrators can require selected GitHub Copilot customizations to come from plugins or managed settings, rather than standalone user or project configuration. The strictPluginOnlyCustomization managed setting can cover all five categories or individual ones: skills, custom agents, hooks, MCP servers, and executable Copilot extensions. Extensions here include executable canvas providers, not VS Code Marketplace extensions. Eligible plugins and managed settings remain allowed. Plugin-only is not a marketplace allowlist. Pair it with strictKnownMarketplaces to restrict marketplace sources. Existing managed settings can supply marketplace definitions and require or block specific plugins. The controls are configured on GitHub Enterprise Cloud for members with Copilot Business or Copilot Enterprise licenses issued by the enterprise or one of its organizations.

    Source for Plugin-only Copilot customizations (opens in a new tab)
  26. VS Code retires the Local harness in February 2027

    On November 16, 2026, GitHub announced that Visual Studio Code 1.146 deprecates the Local harness, the older software that coordinates the model, context, and tools for GitHub Copilot agent sessions in the editor, and that it will be removed in February 2027. The Microsoft Copilot harness, built into VS Code and the default for new sessions since VS Code 1.143, is the replacement. Choosing it does not move files to the cloud. During the three-month transition the Local harness remains available where an organization's policy permits, and new features focus on the Copilot harness. That harness shares its coding-agent runtime with Copilot CLI and the GitHub Copilot app, can continue a session between the editor and the Agents window while the Agent Host stays running, and can enforce managed runtime permissions and sandbox restrictions the Local harness cannot. Selecting the Copilot harness does not turn sandboxing on. Administrators should review any vscode.temporaryDefaultToLocalHarness exception before February 2027, because that setting is removed with the Local harness. To switch, start a new session and select Copilot under Session Target.

    Source for VS Code retires the Local harness in February 2027 (opens in a new tab)
  1. langchain-openai 1.8.3

    Released November 19, 2026. langchain-openai 1.8.3 enables Codex prompt caching on GPT-6.2.

    Source for langchain-openai 1.8.3 (opens in a new tab)
  2. langchain 1.5.1

    Released November 18, 2026. langchain 1.5.1 keeps warning suppression local to the current context and avoids warning-filter races in provider inference and ModelRequest.

    Source for langchain 1.5.1 (opens in a new tab)
  3. langchain 1.5.2

    Released November 18, 2026. langchain 1.5.2 tags synthetic human-in-the-loop tool messages with source metadata.

    Source for langchain 1.5.2 (opens in a new tab)
  4. langchain-anthropic 1.7.11

    Released November 18, 2026. langchain-anthropic 1.7.11 supports Claude Fable 5.2 and Claude Mythos 5.2, sends langsmith_gateway headers from the client, emits a chunk when an empty thinking block starts, raises ModelRefusalError for refusals when tool calls are not forced, and preserves nested caller on server calls and results. Partner packages drop Python 3.10.

    Source for langchain-anthropic 1.7.11 (opens in a new tab)
  5. langchain-core 1.8.3

    Released November 18, 2026. langchain-core 1.8.3 keeps warning suppression local to the current context.

    Source for langchain-core 1.8.3 (opens in a new tab)
  6. langchain-deepseek 1.2.1

    Released November 18, 2026. langchain-deepseek 1.2.1 sends langsmith_gateway headers from the client.

    Source for langchain-deepseek 1.2.1 (opens in a new tab)
  7. langchain-groq 1.2.1

    Released November 18, 2026. langchain-groq 1.2.1 sends langsmith_gateway headers from the client.

    Source for langchain-groq 1.2.1 (opens in a new tab)
  8. langchain-openai 1.8.2

    Released November 18, 2026. langchain-openai 1.8.2 supports GPT-6.2 Astra and Sol, sends langsmith_gateway headers from the client, and stops sharing one async HTTP client between models.

    Source for langchain-openai 1.8.2 (opens in a new tab)
  9. langchain-xai 1.4.1

    Released November 18, 2026. langchain-xai 1.4.1 sends langsmith_gateway headers from the client.

    Source for langchain-xai 1.4.1 (opens in a new tab)
  10. langchain-core 1.8.2

    Released November 17, 2026. langchain-core 1.8.2 adds headers to langsmith_gateway and ignores proxies. The docs clarify when to override aparse_result. Schema creation avoids warning-filter races.

    Source for langchain-core 1.8.2 (opens in a new tab)
  11. langchain-openai 1.8.1

    Released November 17, 2026. langchain-openai 1.8.1 recovers function-call arguments from done events.

    Source for langchain-openai 1.8.1 (opens in a new tab)
  1. Gemini 4 Neon on AI Gateway

    On November 18, 2026, Google's Gemini 4 Neon became available on Vercel AI Gateway. Vercel describes it as a major upgrade over the earlier, lighter Gemini 4 model, with output quality close to Google's larger Gemini 4 models and faster responses, built for coding, agent workflows, and multi-step tasks. It accepts text, images, video, and documents, with a 1M-token context window. It supports tool calling, structured outputs, grounding with Google Search and Google Maps, and four reasoning levels: low, medium, high, and xhigh. The model ID is google/gemini-4-neon. Call it through the AI SDK, the Chat Completions API, the fx coding agent, or Claude Code after npx vercel@latest ai-gateway setup. Coding agents connected to AI Gateway, including Codex and Cursor, can select it from the coding agents guide.

    Source for Gemini 4 Neon on AI Gateway (opens in a new tab)
  2. Parallel web search free on AI Gateway through January 31

    On November 18, 2026, Parallel web search in fast mode became free on Vercel AI Gateway through January 31, 2027, and Parallel became the default web search for eve agents. The tool works with any AI Gateway model, with no separate Parallel API key. It returns links with short excerpts, and results can be filtered by website or publication date. In eve, the built-in web_search tool uses Parallel fast mode automatically when the model is served through AI Gateway. In the AI SDK, add gateway.tools.parallelSearch({ mode: 'fast' }) to the tools. The offer covers fast mode only, with Zero Data Retention enabled by default. Other modes are billed at standard pricing.

    Source for Parallel web search free on AI Gateway through January 31 (opens in a new tab)
  3. Decision-1.1 on AI Gateway

    On November 17, 2026, Decision-1.1 from Microsoft AI became available on Vercel AI Gateway. Vercel says it is more accurate, better calibrated, and faster than Decision-1. It returns structured answers instead of free-form text: Boolean is the probability that a condition is true, Choice is a selected category and probabilities for each option, and Score is a numerical rating on a scale you define. Multiple questions can share the same input in one request. The request format matches Decision-1. The model ID is microsoft/decision-1.1. Call it through the AI SDK with experimental_decide, the OpenAI-compatible Decisions API, or the AI CLI. The TypeScript examples require ai 7.0.128 or later, and the Python examples require the OpenAI Python SDK 3.26.0 or later. Decision requests appear in request logs, count toward budgets, and support decision fallbacks.

    Source for Decision-1.1 on AI Gateway (opens in a new tab)
  4. GPT-6.2 Astra and Sol on AI Gateway

    On November 17, 2026, OpenAI's GPT-6.2 Astra and GPT-6.2 Sol became available on Vercel AI Gateway. Vercel describes Astra as the model for complex, multi-step projects, with improvements to implementing and testing changes in existing codebases, following document templates, and computer use. Vercel describes Sol as the model for everyday coding, writing, and agent workflows, including building interactive web applications from a written brief. The model IDs are openai/gpt-6.2-astra and openai/gpt-6.2-sol. Call them through the AI SDK, the Chat Completions API, or the Responses API. Both support fast mode and OpenAI's Ultrafast service tier. Set providerOptions.openai.serviceTier to ultrafast in the AI SDK, or service_tier to ultrafast in Chat Completions and Responses. Vercel says GPT-6.2 Astra's Ultrafast output-token price is 60% lower than GPT-6 Astra's. Coding agents connected to AI Gateway, including Codex and Cursor, can select either model after npx vercel ai-gateway setup.

    Source for GPT-6.2 Astra and Sol on AI Gateway (opens in a new tab)
  5. MAI-Image-2.7 on AI Gateway

    On November 17, 2026, Microsoft's MAI-Image-2.7 and MAI-Image-2.7-Flash became available in public preview on Vercel AI Gateway. Both raise the maximum output size to 2048 by 2048 pixels and add transparent backgrounds for generated and edited images. Generate or edit images with one AI Gateway API key, without a Microsoft account. Vercel says both support photorealistic images, text rendering, and targeted edits, and up to five reference images. microsoft/mai-image-2.7 is for final assets where image quality is the priority. microsoft/mai-image-2.7-flash is for faster, lower-cost generation when exploring prompts or creating variations. Use either model ID with AI SDK 6 or later. Set providerOptions.azure.background to transparent for a transparent PNG. The AI CLI can generate and edit with the same key and model IDs.

    Source for MAI-Image-2.7 on AI Gateway (opens in a new tab)
  6. Hy4 Flash on AI Gateway

    On November 16, 2026, Tencent's Hy4 Flash became available on Vercel AI Gateway, with an introductory 50% discount on input, cached input, and output tokens through November 29, 2026. It is the lower-cost model in the Hy4 family and keeps the Hy4 feature set: reasoning with preserved thinking across turns, tool calling, structured outputs, prompt caching, and image input, with a 1M-token context window. Reasoning can be turned off. The model name is tencent/hy4-flash. Call it through the AI SDK, fx, the Chat Completions API, or a coding agent connected to AI Gateway, including Claude Code, Codex, and Cursor after npx vercel ai-gateway setup.

    Source for Hy4 Flash on AI Gateway (opens in a new tab)
  7. Private Inference on AI Gateway

    On November 16, 2026, Vercel AI Gateway added Private Inference, which connects a team's own model endpoints to AI Gateway. Fine-tuned, self-hosted, or prerelease models can use the same authentication, SDKs and APIs, and observability as catalog models. Only that Vercel team can view or call the private models, and they are not listed in the public catalog. Credentials are stored encrypted, and Zero Data Retention can be enabled per provider. Requests use a private/ model ID through Chat Completions, Responses, or Messages, and AI Gateway translates each request to the endpoint's format. An optional fallback can call a catalog model if the private endpoint fails. Coding agents connected to AI Gateway, including Claude Code, Codex, and Cursor, can select a private model after npx vercel ai-gateway setup. Private requests appear in logs, trace drains, and custom reporting next to catalog traffic. Providers and models can also be registered with the Vercel CLI, version 65.16.0 or later.

    Source for Private Inference on AI Gateway (opens in a new tab)
  8. Seedance 2.5 Flash on AI Gateway

    On November 16, 2026, ByteDance's Seedance 2.5 Flash became available on Vercel AI Gateway. It is a faster, lower-cost version of Seedance 2.5 for generating video with synchronized audio from a text prompt or from image, video, and audio references. Clips can be 4 to 30 seconds at up to 720p and 24 frames per second, in landscape, square, or portrait. Existing videos can be edited or extended, and images can set the first and last frames. The model name is bytedance/seedance-2.5-flash. AI SDK 7 calls it with experimental_generateVideo. Reference images and videos are passed in inputReferences and named in the prompt as [Image 1] and [Video 1].

    Source for Seedance 2.5 Flash on AI Gateway (opens in a new tab)
  1. Vercel Board for agent coordination

    On November 18, 2026, Vercel Board became available on all plans. Vercel describes it as an experimental message board that gives agents a shared, secure place to coordinate tasks and keep context together. Each task's discussion and results stay in one thread. Channels control which agents and users can access them and what they can do. Messages and replies sync both ways with Slack, GitHub, and other tools through Vercel Connect. Agents use their own harnesses and execution environments and reach Board through the Vercel CLI or the JavaScript SDK. Install the Board skill with npx skills add vercel/board. vercel board whoami reports the board and identity in use.

    Source for Vercel Board for agent coordination (opens in a new tab)
  2. Pull-request-only preview builds

    On November 17, 2026, Vercel added an Ignored Build Step preset that skips preview builds for branches without an open pull request. Production deployments, and previews for branches that have an open pull request, still build. On GitHub and GitLab projects, skipped preview deployments are canceled before they consume Build CPU Minutes. Open the project's Deployments, Configuration. Under Ignored Build Step, select Only build previews for pull requests from the Behavior menu and save. The preset fills in the ignore command.

    Source for Pull-request-only preview builds (opens in a new tab)
  3. Speed Insights measures soft navigations

    On November 17, 2026, Vercel Speed Insights started measuring soft navigations, client-side route changes that update the page without a full reload. It captures LCP, FCP, INP, and CLS for those navigations and includes them in the Real Experience Score. TTFB is not captured because there is no new document request. Tracking is on by default for projects using Speed Insights, in Chromium-based browsers, for frameworks such as Next.js, SvelteKit, and Nuxt. The Speed Insights dashboard adds a Navigation Type filter for All, Hard, or Soft. Soft-navigation events count toward Speed Insights Plus usage. Existing Plus customers receive an email about when billing for these events begins. Projects that newly enable Plus are billed for them from the start. Set softNavs to false and redeploy to turn tracking off. That option requires @vercel/speed-insights 2.1.0 or later.

    Source for Speed Insights measures soft navigations (opens in a new tab)
  4. Standard builds are usage-based for new teams

    On November 17, 2026, new Pro and Enterprise teams on Vercel no longer include a Standard build slot. Standard builds use the same usage-based billing as other build machines. On Pro, Standard builds cost $0.014 per build minute, and the monthly usage credit applies to build usage. Enterprise rates follow the contract. Standard builds remain billable after switching from Elastic to Standard for teams covered by this change. Select a build machine in team settings or project settings.

    Source for Standard builds are usage-based for new teams (opens in a new tab)
  5. Vercel Blob TTLs

    On November 17, 2026, Vercel Blob could delete a blob automatically. Pass ttlDays when you write a blob and it is deleted that many days after upload, from 1 to 730. The TTL counts from the upload date. ttlDays works on put(), copy(), putImage(), multipart uploads, and presigned upload URLs. Every write result and head() response includes expiresAt. setTtl() changes or removes the TTL of existing blobs, one URL or up to 1,000 at a time, without re-uploading them. Vercel CLI 65.16.0 or later accepts --ttl-days on vercel blob put, put-image, copy, and presign, and vercel blob set-ttl updates existing blobs. Setting a TTL costs one Advanced Operation per started 60 days, on top of the write. A 7-day TTL costs one, and the 730-day maximum costs 13. Clearing a TTL and the deletions themselves are free. This requires @vercel/blob 2.9.0.

    Source for Vercel Blob TTLs (opens in a new tab)
  1. Claude Code 2.1.328

    Released November 19, 2026. Auto mode is available for Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry, and the Claude apps gateway. Interactive sessions with no permission mode start in it. Plan mode and Bash prompts for those models on third-party providers use and offer auto mode, as on newer models. disableAutoMode turns both off. mcp_tool hooks can set onFailure to block, so a hook whose call fails blocks the action, as command and HTTP hooks already can. A warning says when a message from a session that skips permission prompts may be held and dropped by a cloud session. /status and claude doctor note when an OpenTelemetry exporter has no endpoint and its data would likely be lost. CLAUDE_CODE_DISABLE_DEFERRED_TOOL_RESUME lets a host stop claude -p --resume and --continue from running a tool call that a hook deferred. Apps that embed Claude Code can switch on the policy plugin's allowManagedModsOnly option. An admin's managed settings still win. The Claude apps gateway warns at startup when a policy's maxEffortLevel is not a valid effort level. /code-review on most models reports only bugs and CLAUDE.md violations. Run /simplify for reuse, simplification, and efficiency cleanups. Findings are marked Important, Pre-existing, or Nit, and each summary ends with a suggested fix. A clean review's list reads "Code review found no issues." Code Mode no longer runs UserPromptSubmit hooks, managed ones included, on scripts' output. To redact that output, use a PostToolUse hook with updatedToolOutput. /bug asks Yes or No before sending. deniedMcpServers entries of the form claude.ai <name> apply to that connector in cloud and self-hosted runner sessions too. Managed PostToolUse hooks no longer fire for calls refused before they ran. The release also includes a large fix sweep for permissions, Code Mode, MCP, Remote Control, cloud sessions, the self-hosted runner, VS Code, and the terminal UI.

    Source for Claude Code 2.1.328 (opens in a new tab)
  2. Claude Code 2.1.327

    Released November 17, 2026. /install-github-app asks before it changes a repository. claude mcp add-json - reads the server JSON from stdin so a token can stay off the command line, and that form is not available inside a Claude Code session. A resumed session says when earlier messages cannot be loaded. /status lists broken agent, skill, command, and output style files, keybinding errors, an unavailable sandbox, ignored permission settings, and hook matchers that never match. /config notes when other settings override a row. Gateway hint headers include x-claude-code-managed-settings-checksum where managed settings set CLAUDE_CODE_GATEWAY_HINT_HEADERS=1. attributionProject names a session's billing group for Claude apps gateways that set admin.billing_group_tag. /tui classic is the name for the non-fullscreen renderer, and /tui default still works. In VS Code, Claude Code: Add to Chat on the Explorer right-click menu @-mentions the selected files and folders. ExitWorktree asks before removing a worktree that may hold unsaved work, and refuses when nobody can answer unless an allow rule, auto mode, or bypass permissions allows it. claude mcp add --scope project stops, instead of overwriting the file, when .mcp.json lists its servers at the top level with no mcpServers key. Path checks refuse some symbolic links, and some paths now ask for permission. permissions.blockReadsOutsideWorkingDirectories asks about more Bash commands in bypass and auto modes, denies some in auto mode, and refuses more Bash commands that name a path under a Read or Edit deny rule. Auto mode's notice says an upcoming release removes the billed classifier fallback, and auto mode ignores more allow rules that run any package, such as Bash(pnpm dlx *). Bash commands whose variables expand to very long text are denied. A subagent resumed by a sibling's message follows the session's permission settings. Code Mode no longer refuses to start when an installed mod hooks tool calls. A subagent launched without a model stays on the model it started on when resumed, including after /model. The first subagent on a model waits up to 1.5 seconds, or 10 seconds in claude -p, for that model's settings from Anthropic. A session's first Bash command can run shell start-up files before the command is approved. Set CLAUDE_CODE_DISABLE_SHELL_HEAD_START=1 to turn that off. The release also includes a large fix sweep for permissions, the sandbox, MCP, sessions, artifacts, Remote Control, cloud sessions, VS Code, and the terminal UI.

    Source for Claude Code 2.1.327 (opens in a new tab)
  3. Claude Code 2.1.326

    Released November 16, 2026. Code Mode can stay on while agent teams are enabled, and a session with teams on starts in Code Mode if you already turned it on. Startup warns when an --allowedTools, --disallowedTools, or --tools value has an entry that does not look like a complete rule. A message explains why MCP servers stop loading when an administrator adds managed-mcp.json during a session. /status adds a row when the provider, or an environment variable such as DISABLE_TELEMETRY, turns features off. /plugin and claude plugin list note when a malformed plugin.json field such as author or homepage is ignored. Ctrl+O shows hidden startup notices, the empty prompt hints at what to do after a failed request, and a subagent view shows an esc to stop hint where Esc would stop it. OTEL_RESOURCE_INCLUDE_HOST, off by default, adds host.name and host.id to OpenTelemetry resource attributes. Background sessions that lost the LLM gateway the shell exported stop before any model request. The background service no longer ends your own processes, such as an SSH login, when a process number on its list was reused. Rules added in /permissions or at a tool prompt, and /add-dir folders, warn and name the settings file when it could not be written, instead of looking saved. The subagent model set in managed settings now applies in Claude Desktop's Code tab. Compaction on Bedrock, Vertex, and Foundry, and with telemetry off, summarizes instead of sometimes replying to one of your messages. Image-heavy sessions on Vertex AI compact and retry instead of sticking on a 400 error about a 30 MB request limit. When max_tokens exceeds the model's limit, Claude Code retries with the maximum the Anthropic API reports. Auto mode no longer stays on its local classifier, which counts toward usage, because of a stale saved value. /model shows Default as the plan's default model, and choosing it switches back to that default. The release also includes a large fix sweep for permissions, the sandbox, MCP, sessions, Remote Control, sign-in, and the terminal UI.

    Source for Claude Code 2.1.326 (opens in a new tab)
  1. Custom CA and client certificates for Workers VPC

    On November 18, 2026, Cloudflare Workers VPC services could verify private services that use a certificate from your own certificate authority, and authenticate to services that require mutual TLS. Previously, Workers VPC trusted publicly trusted CAs and Cloudflare Origin CA only, so a private CA meant weakening verification with verify_ca or disabled, and origins that required a client certificate were unreachable. Upload the CA certificate and the client certificate with wrangler cert upload, then reference them from the TLS settings of a VPC service. You can use a custom CA, a client certificate, or both. The settings apply to HTTPS connections from http services and to TLS connections from tcp services, including PostgreSQL databases used with Hyperdrive, such as Amazon RDS instances that use the RDS CA bundle. Configuration works with Wrangler 4.164.0 or later, the API, or Terraform provider 5.29.0 or later.

    Source for Custom CA and client certificates for Workers VPC (opens in a new tab)
  2. File attachments in Agent Lee

    On November 18, 2026, Cloudflare Agent Lee could take files with a message. Attach a screenshot, a PDF, or a log, HAR, or configuration file with the paperclip, a paste, or a drag onto the chat, and Agent Lee reads it as part of the question. Each message can include up to four files, up to 14 MiB in total after trimming. PNG, JPEG, and WebP images can be up to 3.5 MiB, PDFs up to 5 MiB, and text files up to 100 MiB. Agent Lee reads images directly, converts PDFs to text and reads them from the start, and reads text files in full or, for long files, from the start and the end. A text file over 1 MiB, or 2 MiB for JSON, HAR, YAML, TOML, and XML, keeps its first and last lines. HAR files are compacted in the browser: request and response bodies and cookies are removed, and credential headers such as Authorization and Cookie are replaced with [redacted]. A PDF that is only scanned images, with no text layer, cannot be read. Older attachments age out of the stored conversation. Agent Lee remains in beta.

    Source for File attachments in Agent Lee (opens in a new tab)
  3. AI Search file summaries

    On November 17, 2026, Cloudflare AI Search could summarize files while indexing them. For each text document that produces at least three chunks, it uses a large language model to generate a summary, topics, and keywords. The summary and keywords are stored in the item's context metadata, and a context value you already set is kept. Topics are added to the context header of the file's chunks. File summaries are off by default. Turn them on in the dashboard, or set indexing_options.file_summary to true when you create or update an instance. Changing the setting triggers a full reindex. indexing_options is replaced as a whole on every update, so include the current value of every option. The model calls are billed as Workers AI usage through the instance's AI Gateway, or the default gateway if none is set, and are not part of the AI Search ingestion allotment. Each new or modified file is summarized once.

    Source for AI Search file summaries (opens in a new tab)
  1. Devin Desktop 3.14.30825

    Released November 18, 2026. Password prompts in the command palette, such as the Remote SSH password prompt, hide what you type. The Plan Editor no longer opens plan files as a blank page. Devin Local no longer fails to sign in with "not authorized to use the CLI" for accounts that have Devin Desktop access without CLI access. In remote windows, Open Folder and Add Folder to Workspace accept a completed folder path when you press Enter or click OK or Add.

    Source for Devin Desktop 3.14.30825 (opens in a new tab)
  2. Devin Desktop 3.13.4

    Released November 17, 2026. The Agent Command Center adds tabs, so several sessions and pages can stay open in one window, each with its own Back and Forward history. Local sessions run in a background process, so closing or reopening Devin Desktop or the CLI does not stop a session, and the same session can be open in more than one client. The Files tab previews local PDFs, shows CSV and TSV files as tables, and plays more media types. Devin Local nests each subagent's output under that subagent and shows how many are running. The embedded IDE saves a file on Cmd/Ctrl+S instead of automatically, and a dot marks unsaved tabs. Chat diffs and line counts use Devin's diff colors instead of the editor theme's. New sessions start on the agent starred in Settings > Agents, or on the agent picked last. A Devin Local session keeps accepting messages after its tab is closed and reopened. Devin Local no longer disconnects shortly after launch when another agent has an update, and the first message of a new Devin Local session no longer fails with "ACP host has not authenticated". A command that writes many large files no longer crashes the window. Sign-in no longer fails with "Problem with login service" on networks that reset some connections. Crashes of Devin Desktop windows on some Linux systems, and of the Agent Command Center after some updates on Windows, are fixed.

    Source for Devin Desktop 3.13.4 (opens in a new tab)
  3. Devin Desktop 3.14.30819

    Released November 17, 2026. Devin Bots are available: give a Bot an outcome to own in its chat, and it plans the work and runs sessions until the work is done. Enterprise administrators can turn on proxy auto-detection for the Devin Desktop language server on managed macOS, Windows, and Linux machines with the DetectProxy device policy or the devin.network.detectProxy setting. The image viewer includes zoom buttons and supports Cmd/Ctrl with +, -, and 0. Repeated git push commands to the same branch and commit in a session's worklog collapse into one row. Long streaming replies use less memory when several sessions stream at once. Devin Local keeps images in a conversation until the model provider's request limit is reached, instead of removing them after 20 images. Hidden background session tabs do not start an embedded VS Code instance until you switch to them. Updating Devin Desktop on Windows closes a running window before the update starts and no longer leaves an old Devin.exe in the install folder. In IDE mode, picking a page element or console output in a Devin Local session's browser preview adds it to the chat composer. Scrolling, moving focus, and copying in a session or Devin Review page no longer act on a background tab when the same page is open in more than one tab. Turning Use my ChatGPT plan on or off updates the model picker in open Devin Local chats immediately. Links inside cards in the Agent Command Center that previously ignored clicks open in Devin Desktop or in the browser. Playbooks that are shared for use only no longer show Edit or Delete buttons that fail when clicked. On macOS, switching from the Agents view to the IDE with an IDE banner enabled returns the window traffic-light buttons to the title bar. Devin Local retries a stalled model stream after 40 seconds instead of staying on Running tools for 4 minutes. Long Devin Local sessions on Anthropic models preserve the prompt cache across turns.

    Source for Devin Desktop 3.14.30819 (opens in a new tab)
  1. Gemini 4 Neon generally available

    On November 18, 2026, Gemini 4 Neon (gemini-4-neon) became generally available in the Gemini API. Google describes it as frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows. It is available on the paid tier at $3.00 per 1M input tokens and $15.00 per 1M output tokens. It supports four thinking levels, low, medium, the default high, and xhigh, a 1M-token context window, and long decode up to 1M tokens.

    Source for Gemini 4 Neon generally available (opens in a new tab)
  2. Office files and media slicing in the Gemini API

    On November 18, 2026, the Gemini API could take Microsoft Office and OpenDocument files, including Word, PowerPoint, and Excel, directly. Prompts and function responses can also slice documents, spreadsheets, videos, and audio. A request can select page ranges, worksheet tabs, cell ranges, time intervals, or content layers such as page images, extracted text, formulas, or video transcripts, so only that content is processed and billed.

    Source for Office files and media slicing in the Gemini API (opens in a new tab)
  3. Voice editing in Gemini 3.5 Transcribe

    On November 17, 2026, gemini-3.5-transcribe and gemini-3.5-transcribe-live added voice-driven editing, rewriting, and formatting of dictated speech or pre-existing text passed as source_text.

    Source for Voice editing in Gemini 3.5 Transcribe (opens in a new tab)
  1. The new Netlify experience

    On November 18, 2026, Netlify started rolling out a redesigned experience for building with agents and with a team. You can write or dictate a prompt and choose an agent and model, or publish by dragging files, importing from Git, connecting an agent you already use, or running netlify deploy. Git pushes and agent prompts land in the same app. Every change creates a version with a preview that can be reviewed, archived, or published, including projects that are not connected to Git. Under Manage, Your app, Who can see it, a site can require a password or a Netlify team login. Connectors from the November 17 launch work here, so an agent can use a connected account such as Notion while it builds. The new experience starts with most new users. Existing users can join a waitlist. Connectors remain available on credit-based plans, including Enterprise plans that use credits.

    Source for The new Netlify experience (opens in a new tab)
  2. Agent Runners connectors

    On November 17, 2026, Netlify Agent Runners could use connectors for third-party services. Ask the agent to use a service and it helps set up the connector, for example building from a Linear issue or a Notion spec, pulling data from Snowflake, or setting up a Stripe checkout. If a feature needs an ongoing service, the agent asks for credentials and saves them as a project environment variable. Netlify says more than 150 connectors are available, including Algolia, Amplitude, Databricks, Datadog, Snowflake, Clay, Slack, Zapier, HubSpot, Salesforce, PayPal, Plaid, Stripe, GitHub, Jira, Linear, and Notion. Each person connects with their own account. Other team members cannot use or view that connection. Connectors are available to anyone who can use Agent Runners on a Credit plan or an Enterprise plan that uses credits.

    Source for Agent Runners connectors (opens in a new tab)
  3. GPT-6.2 Astra and Sol in AI Gateway and Agent Runners

    On November 17, 2026, OpenAI's GPT-6.2 Astra and GPT-6.2 Sol became available through Netlify's AI Gateway and Agent Runners with no separate API key. Netlify Functions can call them with the OpenAI SDK and the Responses API. The model ids in the changelog examples are gpt-6.2-astra and gpt-6.2-sol. They are also available in Scheduled Functions, Background Functions, and Edge Functions, and those calls use Netlify's caching, rate limiting, and authentication.

    Source for GPT-6.2 Astra and Sol in AI Gateway and Agent Runners (opens in a new tab)
  1. OpenCode 2.3.1

    Released November 18, 2026. A shell command still running when its timeout expires moves to the background instead of being killed. The shell_timeout_background flag is gone, and a foreground timeout of 0 now means the default 120 seconds. Background compaction is on by default: as a session nears automatic compaction, the summary is written in the background so the next prompt does not wait. Set compaction.background to false to turn it off. A background summary that is never used still costs tokens, and experimental.background_compaction is removed. GPT 6.2 Sol and GPT 6.2 Astra, plus their fast variants, show up in the model picker when you sign in with ChatGPT. Settings, Extensions can check for plugin updates and install one. Selecting code in diff and file viewers selects text so you can copy it, and an Add comment toolbar appears beside the selection. OpenCode again reads CLAUDE.md in any directory that has no AGENTS.md, and falls back to ~/.claude/CLAUDE.md when the global AGENTS.md is missing. OPENCODE_DISABLE_CLAUDE_CODE or OPENCODE_DISABLE_CLAUDE_CODE_PROMPT skips every CLAUDE.md. websearch returns a flat list of results, each with url, title, content, and published date. Code Mode and plugins get a structured read result, Code Mode programs can use Node's Buffer, search("issues") is shorthand for a query object, and tool summaries in the catalog can be up to 500 characters. A malformed filename glob is reported as an error instead of returning no results. Server crashes return a 500 with a JSON UnknownError body. Experimental schedules are now automations, behind the automations flag. The schedules flag still works. An automation runs on a time, an interval, a cron, or a pushed MCP event. Active or paused plain prompt schedules carry over. Schedules with a condition or a run limit, and finished ones, are dropped. The schedule HTTP and SDK endpoints are removed. The new icon set is the default. Collapse long messages is a Timeline setting. The desktop app no longer fails to start when drafts.sqlite is a symlink to or a copy of the CLI database: it moves that file aside and creates its own. An app that embeds the SDK with an in-memory database no longer imports legacy auth.json credentials. The release also includes a large fix sweep for the core, the terminal, the desktop app, the SDK, and plugins.

    Source for OpenCode 2.3.1 (opens in a new tab)
  2. OpenCode 2.3.2

    Released November 18, 2026. GPT-6.2 models do not cache a system prompt sent in the Responses instructions field, so every step rebilled it as uncached input. OpenCode now sends that prompt as a leading developer message on every Responses route, including OpenAI with an API key or ChatGPT sign-in, GitHub Copilot, Zen, Azure, and OpenRouter, so it is cached from the second step on. On GitHub Copilot gpt-6.2-sol, uncached input per step dropped from about 14,000 tokens to 3. Compaction through /responses/compact reads the same cached prefix. webfetch structured output no longer repeats the url and format it was called with.

    Source for OpenCode 2.3.2 (opens in a new tab)
  3. OpenCode 1.19.6

    Released November 17, 2026. OpenCode 1 offers an upgrade to OpenCode 2.0. OpenCode 2 downloads in the background and installs only after you choose Restart to upgrade, Upgrade on quit, or Later. Sessions, projects, and settings are kept. Plugins made for OpenCode 1 need an update. Sessions with many images no longer send oversized requests: when the images in a session's history pass 25 MiB, the oldest are replaced with a placeholder. A provider error for a request body that is too large starts compaction and a retry instead of failing the turn.

    Source for OpenCode 1.19.6 (opens in a new tab)
  1. Auggie CLI 1.2.1

    Released November 17, 2026. Stopping a daemon session waits for it to exit and forcibly terminates it if graceful shutdown stalls. Comments and trailing commas no longer replay onboarding, and malformed history is backed up and reset with a short warning. Auggie no longer hangs at startup on Windows while installing its search tools. The prompt appears right away, and a failed install shows a warning. Backspace deletes characters in the interactive TUI on Windows consoles, including the VS Code integrated terminal. Daemon sessions start reliably when multiple local users share a host. Local print sessions save and restore --permission and --add-workspace values when resumed.

    Source for Auggie CLI 1.2.1 (opens in a new tab)
  2. Auggie CLI 1.2.2

    Released November 17, 2026. /import-legacy-sessions works on Windows. Unusable Bash installations are skipped so commands can use another available shell.

    Source for Auggie CLI 1.2.2 (opens in a new tab)
  1. Codex CLI 0.170.0

    Released November 18, 2026. The terminal can negotiate WebRTC for realtime voice calls. OpenTelemetry HTTP exporters can refresh authentication headers automatically. Experimental support keeps identically named MCP servers separate across environments. Oversized WebSocket requests fall back to HTTP without extra retries. MCP OAuth credential writes are protected from races during cancellation, login, and logout. Idle Cloud conversations reconnect after a transient stream loss, and an active recovery turn is kept when a client reconnects. Windows file replacement, sandbox permission preservation, and runtime registration recovery are fixed. Permission updates and child threads keep existing sandbox restrictions. Duplicate file search results from overlapping roots are removed, and thread search snippets work for queries that contain whitespace. The sandbox documentation explains permission profiles and filesystem rules.

    Source for Codex CLI 0.170.0 (opens in a new tab)
  2. Codex CLI 0.169.1

    Released November 17, 2026. The release highlights say none were verified and point at the compare with 0.169.0. That compare enables incremental tools by default for both Amazon Bedrock providers, so tool declarations go out through additional_tools. Set incremental_tools_legacy to false to opt out. Bedrock Mantle commentary is normalized before replay so empty annotations are omitted and the reasoning prefix stays stable. GPT-6.2 Sol and GPT-6.2 Astra are added to the bundled, Amazon Bedrock, and Vertex catalogs, including Bedrock Runtime global and US cross-region variants. GPT-6.2 Sol is the default catalog selection. The GPT-6.1 Sol context window is 500,000 tokens. GPT-6.2 Astra on Bedrock advertises the ultrafast service tier.

    Source for Codex CLI 0.169.1 (opens in a new tab)
  1. CrewAI 1.15.43

    Released November 19, 2026. CrewAI 1.15.43 fixes detection of multimodal support for Claude 5.

    Source for CrewAI 1.15.43 (opens in a new tab)
  2. CrewAI 1.15.42

    Released November 18, 2026. ChromaDB is an optional extra. Knowledge defaults to local Qdrant and storage no longer depends on ChromaDB. Standalone agents can pick tools and add them later. The release also includes a large fix sweep for task outputs, tool names, Flow IDs, and knowledge stores.

    Source for CrewAI 1.15.42 (opens in a new tab)
  1. Generate multiple images from one prompt

    On November 18, 2026, Figma Design, FigJam, Figma Slides, Figma Draw, and Assets could generate more than one image from a single Make image or Edit image prompt. One model can return two, four, or six images. The same prompt can also run across up to three image models at once, one image from each, so the results can be compared side by side.

    Source for Generate multiple images from one prompt (opens in a new tab)
  2. Allowlist for OAuth apps and MCP clients

    On November 17, 2026, Figma Organization and Enterprise admins could allowlist approved OAuth apps, including MCP clients such as Cursor and Claude Code, to manage which apps their teams can use. Admins can review app requests and manage approved apps in the new Extensions tab in the admin console. Figma says this is rolling out to Organization and Enterprise plans over the coming days.

    Source for Allowlist for OAuth apps and MCP clients (opens in a new tab)
  1. langgraph 1.2.23

    Released November 19, 2026. langgraph 1.2.23 gives Send-dispatched tool calls the same state as an edge, and avoids warning races in dynamic schemas.

    Source for langgraph 1.2.23 (opens in a new tab)
  2. langgraph-prebuilt 1.1.3

    Released November 19, 2026. langgraph-prebuilt 1.1.3 gives Send-dispatched tool calls the same state as an edge, and makes wrap_tool_call state overrides reach the tool.

    Source for langgraph-prebuilt 1.1.3 (opens in a new tab)
  1. Pydantic AI 2.71.0

    Released November 19, 2026. The GitHub release is titled 2026-11-18. GoogleModel supports Gemini 4 Neon (gemini-4-neon). ModelRouter routes agent runs. ToolBridge keeps the tool list fixed behind tool_search, tool_describe, and tool_call so the prompt cache holds for large or changing tool sets. PermissionRule in Harness Permissions can allow, ask about, or deny tool calls and shell commands. Goal keeps unattended runs working until a caller-defined completion check passes. RequireVerification sends a coding agent back to run its checks when it tries to finish after an unverified edit. ReportStaleFiles tells the model when workspace files it read change or are deleted. ReportRemainingBudget, ReportToolErrors, WrapUpAtLimits, LabelCrossModelHistory, CurrentTime, and DenyUnanswered are new harness capabilities. Shell(secret_env=) masks secrets it injects into tool results. Peers(network) joins an InMemoryPeerNetwork under the agent's name, and PeerNetwork is renamed PeerConnection. StavProvider and Otari are new gateway providers. Remote CodeMode turns run as Temporal activities. The release also includes a large compatibility and fix sweep for model providers, deferred tools, and Code Mode.

    Source for Pydantic AI 2.71.0 (opens in a new tab)
  2. Pydantic AI 2.70.0

    Released November 17, 2026. Two high-severity fixes ship in pydantic-ai-harness 0.70.0. With strip_ansi set to true, ToolOutputLimits removed ANSI escape sequences with a pattern that took time quadratic in the length of a line, so a tool result of a few hundred kilobytes with many unterminated escape sequences could occupy the event loop for tens of seconds to minutes. The jwt pattern in secret redaction, included by default in redact_secrets and secret_data(), backtracked the same way on crafted text. The fixed pattern no longer matches a JWT whose header segment contains -eyJ. This release forwards service_tier to xAI and records the served tier, lets Skills read a loaded skill's bundled files when read_files is true and substitute ${SKILL_DIR}, lets a SpendLimits Budget resolve its ceiling from the run context, lets PlaywrightBrowser provision its CDP endpoint per run, lets a parent check a running background delegation with check_task and DelegationTasks.snapshot(), adds StreamGuardrail for streamed text, thinking, and tool-call parts, and adds PruneToolResults so the model can replace a tool result it just received with a summary. PostgresMemoryStore prefix filters can use an index. A finished sub-agent task can be resumed from message_task. The release also includes a large compatibility and fix sweep for model providers, file input, and harness tools.

    Source for Pydantic AI 2.70.0 (opens in a new tab)
  1. Zed 1.29.1

    Released November 18, 2026. This week's release includes a Git Panel branch mode that lists every file changed since a base branch, a bookmark picker, configurable Git date formats, Claude Fable 5.2 for Anthropic BYOK users, plus Claude Mythos 5.2 if your organization has access, and a setting to turn off automatic agent thread titles. agent.auto_generate_thread_titles defaults to true. Set it to false to stop automatic titles for Zed Agent threads. Compare Against Base, in the menu at the top of the Changes tab, or git panel: show branch changes, lists files changed since a base branch such as origin/main. Those files open as diffs and cannot be staged or restored. git.date_format sets absolute Git dates to locale, iso, or a strftime format, and date_style settings choose relative or absolute dates in blame, the branch picker, the stash picker, the Git Panel, and the Git Graph. A bookmark picker, opened with project bookmarks: toggle, browses, searches, and previews bookmarks in the workspace. Helix mode again uses the system clipboard for yank, delete, and paste, as it did before 1.28. Set vim.use_system_clipboard to never to keep the 1.28 behavior. The release also includes a large fix sweep for the agent, Git, the file finder, and the terminal.

    Source for Zed 1.29.1 (opens in a new tab)
  2. Zed 1.29.2

    Released November 18, 2026. Zed publishes an official Snap package for Linux.

    Source for Zed 1.29.2 (opens in a new tab)
  1. Linked billing for the Claude API

    On November 17, 2026, Anthropic launched Linked billing in beta with the user-profiles-billing-2026-10-01 beta header. A customer's own Claude organization can pay for the Claude usage in your app. Once the customer links their organization, set billing_policy on each /v1/messages or /v1/chat/completions request to choose who pays. In the Claude Console, the customer can set a spend cap for your app or unlink it. Linked billing works for customers whose organization pays Anthropic directly for API usage.

    Source for Linked billing for the Claude API (opens in a new tab)
  1. Music generation in the AI Gateway

    On November 17, 2026, the Base44 AI Gateway added a music endpoint, POST /audio/music. A backend function can turn a prompt with a style, a mood, and optional lyrics into a song or an instrumental track. The request is synchronous and returns a link to the finished MP3 in the app's storage. Google's Lyria 3.5 is the default model. dry_run estimates the credit cost before music generation is exposed to the app's users.

    Source for Music generation in the AI Gateway (opens in a new tab)
  1. Devin release notes for November 16

    On November 16, 2026, the sidebar shows sessions you are subscribed to. Mark done replaces Archive: it moves a session to the Done folder while the session keeps running and stays in the sidebars of its other subscribers. Devin keeps a phonebook of each member's linked identities across Devin, Slack, Microsoft Teams, git providers, Jira, Linear, and Pylon, and a mention in Devin tags the person in Devin and in Slack at the same time. Typing # in a cloud session composer picks a Slack channel. The Share dialog can import contacts from Slack. Searching a session's events matches each word on its own. Security Findings can be exported to CSV. Security > Configure is renamed Customize, where security-profile plugins are managed, a profile's skills can set scan paths and include or exclude globs, a threat-model rule can be limited to chosen repositories, and skills and rules stay writable when shared uploads are off. The browser tab is named after the open finding, severity is spelled out in the finding title, and the finding sheet crossfades between findings. Enterprise Knowledge managers can read and maintain skills migrated from Knowledge. Customize > Skills lists the layers skills can be written to, and malformed plugin content says what is wrong. Tagging Devin in a Slack thread whose session is too old to resume starts a new session. !ultrafast is an alias of !fastultra in Slack and Teams. On Microsoft Teams, @user mentions are written back as native mentions, a quoted message names the mentioned person, and message headers name Slack or Teams.

    Source for Devin release notes for November 16 (opens in a new tab)
  1. Gemini CLI 0.69.0

    Released November 19, 2026. /stats heatmap shows daily token usage. Custom themes can live in a standalone theme.json. /model validates model names and accepts the model as a direct argument. model.requestTimeoutMs and GEMINI_API_HEADERS_TIMEOUT_MS override the 60 second API headers timeout in every auth mode. Thought-signature-only turns are dropped and multimodal tool responses are kept. Multi-part tool outputs are bounded and older multimodal inline data is evicted. A quota error no longer retries for hours, and ACP clients are told when a retry starts. Loading no longer hangs for minutes on a network failure. Resuming a session no longer sends a model turn that the API rejects, and the main thread no longer freezes while writing JSONL. Headless abort kills child processes, records cancelled tools, and keeps the signal exit code. Folder trust is resolved before the macOS Seatbelt sandbox. Literal $ characters survive template placeholder replacement. The release also includes a large fix sweep for the CLI, ACP, file tools, and Windows.

    Source for Gemini CLI 0.69.0 (opens in a new tab)
  1. Draft test databases, archived chats, and SSO sessions

    On November 16, 2026, a Lovable project that uses the built-in backend gives a draft its own backend, filled with example data. A change that stores new information, such as a new form field, can be tested in the draft without changing the live app. When the draft is accepted, Lovable applies the change to the live database before publish and leaves the test data behind. Chats can be archived from Chats to leave the chat list without being deleted, and found later under the Archived filter. Archiving stops work Lovable is still doing in the chat, including automations that chat runs on a schedule, and Lovable can archive chats it started. Workspace admins and owners on Business and Enterprise plans can set the SSO session duration to 30 days or Indefinite. Indefinite does not ask members to log in through SSO again. A workspace that has not chosen a duration uses Indefinite.

    Source for Draft test databases, archived chats, and SSO sessions (opens in a new tab)
  1. Mastra core 1.84.0

    Released November 18, 2026. The GitHub release is titled November 17, 2026. Agents created with createDurableAgent or createEventedAgent, and evented workflows, can opt into experimental storageVersion 2, which stores a run as a header plus per-step rows so long runs write only deltas. Mastra's example is 98 MB down to 11 MB published, and 75 MB down to 3.1 MB written. result.steps() is async and getState() is standardized. startWorkers() starts a workflow recovery loop that restarts crashed default-engine runs without waiting for another instance to boot, using a leader-election lease. Recovery is broker-aware: on Redis or Valkey Streams and Google Cloud Pub/Sub, instances no longer restart every run on boot and instead let the broker redeliver unacked step messages. knowledge.movePage() moves a page and its descendants without losing ids, embeddings, or links. knowledge.reembed() re-embeds existing pages and notes in resumable batches. @mastra/platform adds hosted Platform Memory and Knowledge clients, Platform Workspace re-exports, and PlatformBrowser for Stagehand sessions backed by Platform browser sessions. The built-in MCP server can render observability results as Studio-based MCP Apps, a Trace Viewer and an Observability Explorer, in clients that support MCP Apps. Compact text goes to the model and the full rows go to the app. Breaking changes: in knowledge alpha scope matching, $param matches exactly one segment and a raw colon is no longer allowed inside a scope segment, so some existing scopes can read as empty until ids are updated. newestOpenPullRequestNumber is no longer exported from @mastra/factory/boards. @mastra/playground-ui removes ListSearch, SearchFieldBlock, SectionCard, and CardHeading. MastraFactory no longer accepts intake.evaluator and throws at startup if it is passed. @mastra/core 1.84.0 also adds collectionExists(name) on Factory storage, supported by LibSQL and Postgres.

    Source for Mastra core 1.84.0 (opens in a new tab)

Updated

Newsletter

Get what shipped each week

One email when rankings change or a tool ships something that matters. No spam, unsubscribe anytime.