AI Tool Updates · September 2026

Gemini Gemini 4 Argon Is Announced, but Only Cyber Defenders Get It for Now

Google announced Gemini 4 Argon, a frontier model for coding, enterprise knowledge work and cyber defense. It rolls out in phases, starting with a limited group of security defenders.

What changes You cannot try it in the Gemini app yet. No date has been announced for general users, so wait for the rollout notice rather than changing your workflow.Source →

Claude Code 2.1.285 Claude Code Can Now Turn Off Web Fetching

Setting the CLAUDE_CODE_DISABLE_WEB_FETCH environment variable turns off the WebFetch tool.

What changes When you're working on internal code that shouldn't make outside requests, you can stop Claude Code from pulling web pages on its own.Source →

Claude Code 2.1.285 Claude Code Opens the Desktop App From Your Terminal

claude --desktop opens the Claude desktop app on the current directory. Add --continue or --resume <id> to hand an existing session over to the desktop app.

What changes When a terminal session outgrows the terminal, you can move it to the desktop app without starting over.Source →

Claude Code 2.1.285 Claude Code Shows Which Plugin Settings Are Missing

claude plugin configure <plugin> lists a plugin's options and flags the ones that are unset. With --values-stdin it saves new values read from standard input.

What changes When a plugin won't work because something isn't configured, you can see exactly what's missing, and scripting the setup gets easier too.Source →

Claude Code 2.1.285 Claude Code Configures Bundled MCP Servers at Install Time

claude plugin install --config now accepts <server>.<key>=<value>, so a bundled .mcpb MCP server can be configured at install time and start without a trip to /plugin → Configure.

What changes Rolling the same plugin out to several machines or teammates now takes one install command, settings included.Source →

Claude Code 2.1.285 Claude Code Fixes /ultrareview Rejecting Some File Names on WSL

On WSL, /ultrareview no longer refuses to upload a checkout on a Linux volume when a changed file's name contains a colon or ends in a dot or a space.

What changes If /ultrareview ever refused to upload on WSL for no obvious reason, you can now run the review without renaming files.Source →

Claude Code 2.1.285 Claude Code Fixes /ultrareview Uploading the Wrong Worktree on Windows

On Windows, /ultrareview no longer uploads a linked worktree of a repository rooted at your home folder in some cases.

What changes If you keep a repository in your home folder and work across worktrees, reviews now pick up what you expect.Source →

Codex rust-v0.159.0 · New Features Codex Lets You Scroll the Transcript While Deciding on a Plan

Codex now lets you scroll the transcript while you decide whether to implement a plan.

What changes When judging a plan Codex proposes, you no longer have to back out to recheck the earlier conversation. It is easier to check the plan against requirements you gave.Source →

Codex rust-v0.159.0 · Bug Fixes Tables Copied from Codex Keep Their Markdown

Copying a selection from the Codex transcript now preserves Markdown tables, formatting and whitespace, and more terminals copy automatically on selection.

What changes If you paste a comparison table from Codex into a doc or issue, you no longer need to repair a broken table first.Source →

Codex rust-v0.159.2 · Bug Fixes Codex Stops Flashing Console Windows on Windows

On Windows, Codex no longer flashes a console window when it launches background processes or sandboxed commands.

What changes If you hand Codex a task and keep working in another window, the constant flicker in the corner of your screen is gone.Source →

Codex rust-v0.159.1 · New Features Codex Switches Its Default Model to GPT-6.1 Sol

GPT-6.1 Sol is now the default in Codex's bundled model catalog, and in the Amazon Bedrock Mantle and Runtime catalogs as well.

What changes Unless you've pinned a model, new sessions will run on GPT-6.1 Sol after you update. If the same request suddenly comes back different, this is the first thing to check.Source →

Codex rust-v0.159.0 · New Features Codex Can Now Change Course Mid-Response

With the opt-in instant_interrupt setting, anything you type while the model is responding, or while a long code-mode call is running, steers Codex right away. It's off by default.

What changes When you can see Codex heading the wrong way, you no longer have to wait it out or kill the whole task. Type a correction and it adjusts on the spot.Source →

Codex rust-v0.159.0 · New Features Codex Gets a Leaner Welcome Screen and In-Session Tips

New sessions open with a compact welcome screen and consistent headers, and Codex now shows the occasional tip during and after a turn.

What changes If you're new to Codex, you'll pick up useful features as you work instead of digging through the docs.Source →

Codex rust-v0.159.0 · New Features Codex Clears Warnings You've Already Reviewed

Closing the warnings viewer now dismisses the warnings you've reviewed. Press k to keep one around for later.

What changes You stop re-reading the same warnings every time, and only the ones you flagged stay on the list.Source →

Claude Code 2.1.284 Claude Code's Default Sonnet Is Now 5.5

Claude Code switched its default Sonnet model to 5.5 on the Anthropic API, with a 1M-token context window and pricing of $2/$10 per million tokens for input/output, plus $0.20 per million for cached reads.

What changes If you haven't pinned a model, your next session will automatically run on 5.5. Anyone leaning on prompt caching for long sessions or large codebases should notice lower cache-read costs.Source →

Claude Code 2.1.284 Auto Mode Gets a "Yes, Just This Once" Option

Auto mode's confirmation prompt for reading a file outside the working directory now offers a "yes, but ask again next time" option.

What changes When you just need Claude Code to peek at one file outside the project folder, you no longer have to choose between always allowing it or blocking it every time, just approve that one instance.Source →

Claude Code 2.1.284 Spend Limits Now Show as Dollar Amounts in /usage

For organizations using the Claude apps gateway, `/usage` and the status line now show actual dollar amounts, like "$271.40 / $500.00 spent this month."

What changes If you've been guessing how close your team's account is to its spend limit, you can now see the actual remaining budget on screen.Source →

Claude Code 2.1.284 Reconnect All Failed MCP Servers at Once

Running `/mcp reconnect all` in the interactive terminal now retries every MCP server that failed to connect or needs authentication, all at once.

What changes When several of your MCP servers drop at once, you no longer have to reconnect them one by one, a single command handles it.Source →

Codex rust-v0.158.0 Codex Now Asks for Fewer Redundant Confirmations

Codex now recognizes system path aliases already covered by an existing permission on macOS, so it stops asking for approval on the same operation twice. It also retries an approval review instead of auto-cancelling a pending action just because new input arrived mid-review.

What changes If you've been annoyed by Codex re-asking for the same approval on repetitive tasks, this update should cut down how often that happens, and you can now check on progress mid-task without accidentally canceling it.Source →

Codex rust-v0.158.0 Codex Defaults to Approving Elevated Commands, Adds MCP Auth Options

Terminal input approval is now on by default for commands that run with elevated permissions. Codex can also connect to MCP servers that require a pre-registered OAuth client secret, and direct exec-server WebSocket connections can be secured with bearer tokens.

What changes Codex will now double-check before running commands that need elevated system rights, cutting the risk of an unintended high-privilege action slipping through. Teams running MCP servers with stricter auth requirements also get a wider set of servers they can now connect to.Source →

Codex rust-v0.158.0 Codex Adds a Transparent-Background Option to Image Editing

Image generation and editing can now explicitly request a transparent background, and edits accept images already attached to the conversation as files.

What changes For assets like icons or logos that need a transparent background, you can now get them directly without a separate background-removal step.Source →

Claude app The Claude App Adds Claude Sonnet 5.5

The Claude app now offers Claude Sonnet 5.5. Anthropic describes it as a faster, lower-cost complement to Opus 5.5.

What changes If you already use Sonnet for everyday tasks, you should notice faster responses starting with your next conversation, with nothing to configure.Source →

Claude Code 2.1.283 The /mcp tool list now shows more at once and scrolls

The /mcp command's tool list now shows more tools per screen and can be scrolled with the page keys or the mouse. Tools your organization has blocked are now marked with a warning icon.

What changes If you have many MCP servers connected and the tool list runs long, you no longer need to re-run the command just to find the one you're looking for.Source →

Claude Code 2.1.283 Images from MCP tools are now saved to a file so Bash and Read can open them

Images returned by MCP tools are now also saved to a file, so tools like Bash and Read can open them.

What changes When you use an MCP tool that returns a screenshot or a chart, it's now easier to reuse that image in a later step.Source →

Claude Code 2.1.283 /tasks now shows status icons and supports mouse and page-key navigation

Each row in /tasks now shows a status icon, and the title and key hints stay visible on screen even with many tasks. The list can also be paged with keys or the mouse wheel and clicks.

What changes When you're running several background tasks at once, you can tell at a glance which ones are done and which are still running instead of scrolling to check.Source →

Notion Notion Adds Classic Art Collections for Page Cover Images

Notion added a new collection of timeless artwork you can pick from when choosing a page cover image.

What changes You can pick a cover image for a meeting note or project doc without worrying about image rights.Source →

Notion Notion Adds Muting for Individual Discussion Threads

You can now mute reply notifications for a specific discussion thread instead of all of them.

What changes You no longer have to mute an entire page just to escape one comment thread you don't care about.Source →

Notion Notion Adds Tabs for Organizing Content Within a Page

Notion now lets you split content within a single page into tabs.

What changes Instead of scrolling through a long document, you can click a tab to jump straight to the section you need.Source →

Notion Notion Adds a Way to Archive Outdated Pages in Bulk

Notion added a way to find and archive outdated, unused pages in bulk to keep a workspace clean.

What changes You can tidy up years-old, untouched pages in one pass without deleting them, keeping workspace search results cleaner.Source →

Claude Code 2.1.283 Claude Code Adds availableModelsMatch to Pin Model Versions Strictly

Claude Code added an availableModelsMatch managed setting. Set to "exact," an availableModels entry allows only the exact version named, so a new release stays blocked until it's explicitly added.

What changes If your team's settings pin a specific model version, a new release won't silently swap in without anyone adding it to the list.Source →

Claude Code 2.1.283 Claude Code Adds deniedModels to Block Specific Models Individually

Claude Code added a deniedModels managed setting that blocks specific models even when availableModels would otherwise allow them.

What changes If you need to block just one model over cost or policy concerns, you don't have to rewrite the whole allow list.Source →

Claude Code 2.1.283 Claude Code Now Names the Path When a Plugin Directory Fails to Load

When a --plugin-dir folder fails to load, Claude Code now names the exact path in its stream-json output.

What changes When one of several plugin directories fails to load, you can immediately tell which one without digging through logs.Source →

Claude Code 2.1.283 Claude Code Fixes a Windows Bug That Let cmd Delete Protected Folders

Fixed a Windows bug where the PowerShell tool let cmd /c commands like rd, rmdir, del, or erase delete drive roots, the home folder, and other locations that Remove-Item normally refuses to touch.

What changes On Windows, letting Claude Code run delete commands is now less likely to wipe out a core system folder by mistake or malfunction.Source →

Claude Code Claude Code adds /doctor prompt-audit to check your CLAUDE.md, skills, and agents

A new /doctor prompt-audit command (also available as /checkup prompt-audit) audits your CLAUDE.md files, skills, and agent configs in one pass.

What changes If your project's CLAUDE.md has been edited across many sessions, this one command cuts down the time it takes to spot stale or conflicting instructions.Source →

Claude Code Claude Code lets you click to expand truncated messages from other sessions

In fullscreen mode, truncated messages from your other sessions can now be clicked to expand and show the full text.

What changes When you're running several sessions at once, you can check what another one was doing by clicking the truncated message instead of switching into that session.Source →

Claude app You can now build and submit your own Claude plugins

A new developer portal lets you submit plugins to the Claude directory, track them through review, and see usage analytics once they're live.

What changes If you've built a tool you want others to use inside Claude, there's now an official channel for that, though it still has to pass review before it's listed.Source →

Codex rust-v0.157.0 Codex Now Shows Fullscreen Transcripts by Default and Adds Shift-Click Selection

Codex CLI now defaults to a fullscreen transcript view, and holding Shift while clicking lets you select a range of text.

What changes Scrolling back through a long session to find or copy an earlier answer now takes fewer steps.Source →

Codex rust-v0.157.0 Codex Adds an 'f' Shortcut to Fork Conversations Open Elsewhere

Pressing f in a conversation open in another app now forks that conversation, keeping drafts and queued prompts intact.

What changes Branching off an experiment without losing work in progress in another window takes fewer steps.Source →

Codex rust-v0.157.0 Codex Enables /import in Remote and Background Sessions

The /import command, previously limited to local sessions, now also works in remote sessions and local background-server sessions.

What changes You no longer need to switch to a local session just to import a file into work running remotely.Source →

Codex rust-v0.157.0 Codex Improves Terminal Rendering for Checkboxes and Equations

Terminal output rendering was improved so Unicode bullets, checkboxes, aligned equations, and optimization notation display more cleanly.

What changes Checklists and equations mixed into answers are less likely to render broken or overlapping in the terminal.Source →

Codex rust-v0.157.0 Codex Respects tmux Mouse Settings in SSH Terminal Sessions

Fixed a bug where automatic split-screen mode ignored tmux mouse settings and blocked Terminal.app's native scrollback over SSH.

What changes Mouse scrolling and selection now work as expected again when using Codex over SSH.Source →

Codex rust-v0.157.0 Codex Adds Retries for File Uploads and Extends the Timeout to Five Minutes

Codex now automatically retries transient file-upload failures and has extended the upload timeout to five minutes.

What changes Large files or slow networks are less likely to force you to restart an upload that gets interrupted.Source →

Codex rust-v0.157.0 Codex Enforces Network Restrictions Across Redirects and Ongoing Connections

Network restrictions now apply across redirects and ongoing HTTP and WebSocket traffic, and an already-open connection is cancelled if a policy change revokes its access.

What changes If you run Codex with network access restricted, this closes a gap that let a redirect slip past those restrictions.Source →

Codex Codex adds GPT-6 Sol and Luna, plus Amazon Bedrock support

Codex now supports the new GPT-6 Sol and Luna models, including through Amazon Bedrock, and existing sessions on older models get a migration prompt to switch.

What changes If you're already using Codex, expect a migration prompt in your next session. Teams already on AWS can now reach the new models through Bedrock without setting up separate access.Source →

Codex Codex auto-starts a background server for interactive sessions

Eligible interactive sessions now start a background server automatically, with recovery options offered if something goes wrong.

What changes You no longer need to manually spin up a server to hand off long-running tasks, so you can start a long job the moment your session opens.Source →

Codex Codex keeps voice conversations alive when you switch threads

Active voice conversations now stay connected when you switch to a different thread, instead of dropping.

What changes If you're talking to Codex by voice and need to check another thread, you won't have to restart the conversation when you come back.Source →

Codex Codex recovers unsent answers into the composer when a turn ends

If a turn ends before you finish answering a question, your unsent reply is now recovered into the composer instead of being lost.

What changes Fewer lost replies from bad timing means you won't have to retype an answer you already wrote.Source →

Notion Notion AI chats can now be shared as a link

Notion added the ability to share a conversation with Notion AI as a link others can open.

What changes Instead of screenshotting or copy-pasting an AI answer for a coworker, you can just send them the link and they'll see the whole exchange.Source →

Notion Notion desktop now supports voice input

Notion's desktop app added voice input for typing text.

What changes You can dictate a thought the moment it occurs to you, instead of pausing to type it out.Source →

Claude Code 2.1.282 Claude Code lets you cap prose width in wide terminals

A new maxProseWidth setting lets you cap how wide Claude's prose stretches in wide terminals, while tables and code blocks still use the full width.

What changes If you keep your terminal wide on an ultrawide monitor, you can fix lines of explanation stretching too far to read comfortably with a single setting.Source →

Claude Code 2.1.282 Claude Code now flags disabled telemetry settings on startup

If a project's settings files have telemetry variables that were ignored or turned off, Claude Code now lists them in a startup notice and in /status and claude doctor.

What changes If a project's telemetry settings have been customized, you can see what's turned off right in the startup notice or claude doctor, without digging through settings files.Source →

Claude Code 2.1.282 Claude Code fixed missing web search for new models on Vertex AI

Fixed web search not being offered on Vertex AI for models Claude Code doesn't recognize yet, such as newly released ones.

What changes If you run Claude Code through Vertex AI and picked a newly released model only to find web search missing, it should work now.Source →

Claude Code 2.1.282 Claude Code speeds up resuming very large sessions

Resuming very large sessions, including ones that were never compacted, is now faster.

What changes When you're running several long sessions, you'll spend less time waiting for a previous one to reload.Source →

Claude Code 2.1.282 Claude Code explains why a session failed to resume on Windows

When a session on Windows can't be resumed because its transcript file couldn't be read (EBADF), the error message now names possible causes and what to try.

What changes When a session won't open on Windows, you won't be stuck staring at a bare error code, the message itself points you toward the cause.Source →

Gemini Gemini launches 3.8 Live with a real-time talking avatar

Google introduced Gemini 3.8 Live along with a Live Avatar feature that shows an animated face reacting and talking back to you in real time.

What changes Adding a reacting face to what used to be a voice-only live conversation pushes the product closer to feeling like a video call with someone actually there.Source →

Claude Code 2.1.281 Claude Code lets you turn off commit and PR attribution with one setting

Adding "attribution": false to settings.json now removes all the Claude attribution that gets added to commits and PRs automatically. Older CLI versions skip the settings file that holds this value, though, so keep it in object form in files shared across versions.

What changes If your team's repo policy doesn't allow tool attribution in commit messages, you no longer have to strip it out by hand every time; one setting handles it.Source →

Claude Code 2.1.281 Claude Code no longer leaves a stuck dialog when an MCP server asks it to open a browser flow

Claude Code added URL-mode elicitation on MCP protocol connections, letting a server ask it to open a browser-based flow, and it no longer leaves a waiting dialog on screen when the server has no way to confirm completion.

What changes If your MCP integration needs a browser-based step like a login, you won't be left staring at a stuck waiting dialog after it's done, which used to mean restarting the session.Source →

Claude Code 2.1.281 Claude Code's plugin validator now catches broken MCP configs before they fail silently

The claude plugin validate command now also flags .mcp.json entries that would be silently dropped at load time, undeclared ${user_config.*} references, and insecure URLs.

What changes Instead of spending time figuring out why an MCP server silently isn't connecting, you can catch the misconfiguration up front with a single validate command.Source →

Claude Code 2.1.281 Claude Code's /insights now estimates how much auto mode would cut down permission prompts

The /insights command now includes an auto mode recommendation that estimates how many permission prompts auto mode could have handled in your recent sessions.

What changes If the constant permission prompts have been annoying, you can now see a concrete estimate of how many auto mode would have saved before deciding whether to turn it on.Source →

Copilot PowerPoint Build Your Own Custom Skills for Copilot in PowerPoint

Microsoft 365 Copilot now supports user-defined custom skills in PowerPoint on the web. Before, only built-in skills were available.

What changes If you repeat the same instructions every time you build a deck, you can set up a command for your own style once and reuse it, within your company's security and compliance limits.Source →

Copilot Microsoft 365 Copilot extensibility Copilot Now Pulls Live Data From Outside Services Without Storing It

Federated connectors for Microsoft 365 Copilot are now generally available. They connect Copilot to popular third-party services and fetch data in real time over the Model Context Protocol (MCP), without storing or indexing it in Microsoft's services. They work in Researcher, Microsoft 365 Chat, and when editing with Copilot in Excel.

What changes When your work is scattered across outside tools, Copilot can look up the latest version on the spot. Access runs under your own identity, so it's worth being deliberate about what you connect.Source →

Copilot Outlook Copilot Agents Now Work Fully in Classic Outlook for Windows

Declarative agents in Copilot Chat in classic Outlook for Windows now support the full set of extensions: plugins, adaptive cards, mail and calendar actions, and enterprise connectors, bringing classic Outlook in line with other Copilot Chat surfaces.

What changes If your company is still on classic Outlook and agent features that worked elsewhere didn't work there, you can now do the same things without leaving your inbox.Source →

Copilot Microsoft 365 Copilot Chat Ask Copilot Directly Inside Microsoft Forms

Microsoft Forms now has a built-in Copilot chat pane for improving forms and quizzes, drafting invitations, and analyzing responses.

What changes Instead of exporting responses to analyze elsewhere, you can now ask for a summary or insights right inside the Forms screen.Source →

Copilot Microsoft 365 Copilot app Copilot Cowork lets you delegate an outcome, not just a task

Copilot Cowork lets you describe the outcome you need; it gathers context from your emails, meetings, chats, files, and business data, works across tools, and returns completed results.

What changes For work that means pulling information from several places, like assembling meeting materials, you can now describe the end result you want in one line instead of listing out each step. It still runs through your approval and your organization's existing security and privacy controls.Source →

Copilot Microsoft 365 Copilot app Copilot's Vision now reads your shared screen and phone camera in real time

Copilot's Vision feature can now analyze desktop screen sharing and live mobile camera feeds in real time. During a voice session, you can turn on your phone's front or back camera and ask about what's in front of you, and Copilot answers out loud using work and web data.

What changes You no longer need to capture and attach a screenshot; you can just point your camera at what's in front of you or share your screen and ask right away, which saves time on things like physical documents or equipment that are awkward to describe in text.Source →

Copilot Microsoft 365 Copilot app The Microsoft 365 Copilot app gets a new lightweight Notebooks workspace

The Microsoft 365 Copilot app now has a new Notebooks screen for organizing related chats, outputs, and references in one place. Copilot grounds its answers in what's accumulated in the notebook, so work continues instead of starting over each time. The fuller-featured version stays in OneNote; this is the quicker, lightweight option.

What changes If you keep coming back to Copilot about the same topic, you can now collect it all in one notebook and pick up where you left off instead of re-explaining the background every time.Source →

Copilot Microsoft Copilot Adds Profile Cards for Rooms and Desks

Buildings, rooms, and desks registered in the Places directory now get profile cards, viewable from Copilot Chat and other entry points.

What changes Before booking a room, you can check its location and amenities right inside Copilot Chat.Source →

Copilot Microsoft Copilot Lets You Add Notebook References from Quick Access in Search

In the Microsoft Copilot app, you can now add a reference to a Copilot Notebook directly from the Quick Access section in Search.

What changes Instead of hunting down a notebook you were using, you can pull it in straight from the search box.Source →

Copilot Microsoft Copilot App Adds Apple CarPlay Support

The Microsoft Copilot app now integrates with Apple CarPlay, letting you use it with voice commands while driving.

What changes While driving, you can check your schedule or ask about an email hands-free.Source →

Copilot Excel Excel shows which edits were made by Copilot in Show Changes

Excel's Show Changes pane now attributes edits and suggestions to Copilot with an on-card label. Previously, Copilot's activity wasn't visible there at all.

What changes In a shared workbook, you can now tell at a glance which cells a person edited versus which ones Copilot touched, making it easier to focus your review on the AI-made changes.Source →

Copilot Planner gets a Copilot agent chat

Microsoft Planner now includes a chat interface for working with a Copilot agent directly inside the app.

What changes You can now ask Copilot for help or hand off tasks without leaving the Planner screen you're already managing your to-dos in.Source →

Copilot Outlook Copilot adds more email triage actions

Copilot in Outlook adds more actions for triaging incoming email.

What changes On a day packed with back-to-back meetings, you have more ways to let Copilot clear out your inbox before you even open it.Source →

Copilot Excel Excel Copilot replies now link straight to the cells they changed

Copilot chat replies in Excel now include links that jump straight to the sheets, tables, ranges, and charts they changed. Previously, you had to read the response and hunt down the change in the worksheet yourself.

What changes You'll spend less time scrolling to find what Copilot actually changed after asking it to edit a spreadsheet. Clicking the link in its reply jumps you straight to the cell.Source →

Copilot Microsoft 365 Copilot Copilot search now shows email attachments right on the result

In Microsoft 365 Copilot search, email results that have attachments now show them as pill-shaped tags right on the result. Previously, you had to open the email to see if it had attachments.

What changes You no longer have to open emails one by one to find an attachment, you can open the file straight from the search results screen.Source →

Copilot Microsoft 365 Copilot Copilot search adds a date sort filter for Outlook mail

Copilot Search added a "Sort by" filter for Outlook Mail results, letting you sort by date in ascending or descending order. Previously, results appeared with no sorting option.

What changes When you need the most recent or the oldest email on a topic, you can sort straight to it instead of scanning through results.Source →

Copilot Microsoft 365 Copilot Copilot now searches chat and transcripts to find meetings

Copilot can now find meetings by topic or keyword using the meeting body, chat messages, and up to two months of transcript text. Previously, meeting search was narrower and didn't include chat or transcript content.

What changes Even if you can't recall a meeting's title, you can find it by searching for a keyword from what was actually said or typed in the chat during it.Source →

Codex rust-v0.156.1 Codex's model catalog now includes GPT-6 Sol and Luna

A 0.156.1 hotfix release added GPT-6 Sol and Luna to Codex's model catalog.

What changes Two new models now show up as choices when you pick a model in Codex, so you can switch between them and compare results for different tasks.Source →

Gemini Gemini chat can now connect directly to 12 outside apps like Airtable and Adobe

Gemini added 12 new Connected Apps you can link directly in chat: productivity tools like Airtable, Linear, and monday.com; creative tools like Adobe, Picsart, and Webflow; and lifestyle services like apartments.com, Experian, and Peloton.

What changes You can dictate meeting notes and push them straight into Linear as tasks, or type "@Peloton" mid-chat to check today's workout plan, all without switching tabs.Source →

Codex rust-v0.156.1 Codex model picker adds GPT-6 Sol and Luna

Codex's model picker now lists OpenAI's new lower-cost GPT-6 Sol and lightweight GPT-6 Luna.

What changes You can route quick, simple tasks to Luna to save on cost and save Sol or your usual model for the harder ones, choosing per session.Source →

Claude Code 2.1.280 Claude Code's default Opus model becomes Opus 5.5

Claude Code's default Opus model is now Claude Opus 5.5, with a 1M-token context window, priced at $4/$20 per million input/output tokens and $0.20 per million cached-read tokens.

What changes If you haven't changed your settings, your next session automatically uses Opus 5.5, and the more your workflow relies on cache reads, the more the cost drops.Source →

Claude Code 2.1.280 Claude Code lets you raise the MCP description length cap

A new environment variable, CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH, lets you adjust the 2,048-character cap on MCP tool descriptions and server instructions for the whole session.

What changes If you've seen a verbose MCP server's instructions get cut off, raising this environment variable stops the truncation.Source →

Claude Code 2.1.280 Claude Code stops deleting needed folders during Windows cleanup

On Windows, fixed a bug where background cleanup deleted the symlink or junction used to relocate folders like ~/.claude/session-env or image-cache, instead of just cleaning their contents.

What changes If you've relocated your session-env or image-cache folder to another drive on Windows, background cleanup will no longer wipe out that link entirely.Source →

Claude Code 2.1.280 Claude Code makes deleting a permission rule harder to do by accident

The /permissions screen now returns focus to the rule list after you view, add, or delete a rule, and the delete-rule and remove-directory confirmation prompts now default to No.

What changes Accidentally hitting Enter is less likely to delete a permission rule or directory now that the default answer is No.Source →

Codex rust-v0.156.0 Codex fixed logins breaking behind system proxies

Codex fixed logins failing to recover when running behind a system proxy, and fixed MCP credentials not refreshing when the OAuth discovery server returned a 503 error.

What changes If you work behind a corporate proxy and kept getting logged out of Codex, you shouldn't have to fight the login screen as often now.Source →

Codex rust-v0.156.0 Codex voice mode no longer drops speech mid-sentence

Codex fixed speech getting dropped during playback pauses or when a burst of incoming audio arrived at once.

What changes When talking to Codex by voice, fewer words should vanish mid-sentence, so you shouldn't need to repeat yourself as often.Source →

Codex rust-v0.156.0 Codex terminal gets 6 new themes plus Mermaid diagrams and inline equations

Codex's terminal now offers six new color themes, and it can render supported Mermaid diagrams and equations directly inside responses.

What changes When a response includes an architecture diagram or an equation, you can check it right in the terminal instead of copying it out to another viewer.Source →

Codex rust-v0.156.0 Codex adds a /daemon command to control its background server

Codex now lets you restart its local background server with the /daemon command, or skip it entirely with the --no-daemon flag.

What changes If the background server misbehaves, you can restart just that with /daemon instead of relaunching Codex entirely, or skip it altogether with --no-daemon for a lighter run.Source →

Codex rust-v0.156.0 Codex now keeps Plan mode and earlier settings intact when you resume a session

Resuming a Codex session now restores Plan mode instead of dropping it, and editing an earlier prompt preserves the thread's identity and settings.

What changes Your plan no longer disappears when you close and reopen a session, so you spend less time re-explaining context each time you pick work back up.Source →

Codex rust-v0.156.0 Codex closes three sandbox isolation gaps

Codex closed sandbox isolation gaps that involved inbound Windows connections, privileged Linux/macOS sockets, and writes through file handles marked read-only on macOS.

What changes Since Codex's safety model depends on code actually staying inside its sandbox, closing these gaps reduces the paths an agent could unintentionally use to reach outside it.Source →

Codex rust-v0.156.0 Codex adds a searchable fullscreen view

Codex CLI added an optional fullscreen mode toggled with /tui, including transcript search, mouse text selection, and right-click copying.

What changes When running several long sessions, you'll spend less time scrolling back to find an earlier instruction.Source →

Codex rust-v0.156.0 Codex turns on voice conversations by default

After updating, Codex turns on voice conversations by default. Press F8 to toggle it, or use /voice to pick detailed settings. Linux and Windows now get bundled audio runtimes, so no separate install is needed.

What changes You may be surprised to find your mic live right after updating. If you don't want voice input, press F8 to turn it off as soon as you update.Source →

Codex rust-v0.156.0 Codex adds a dashboard for token usage

The new /usage command opens an analytics dashboard showing account usage, total tokens, and plugin/skill activity in one screen.

What changes Checking how many tokens you've used this month no longer means hopping between screens, just run /usage.Source →

Codex rust-v0.156.0 Codex enables worktree sessions by default

You can now filter tasks by status and spin up worktree sessions directly from the agent command center, and worktree support, which lets you run multiple tasks in parallel, is now on by default.

What changes When running multiple tasks at once, you can spin up a per-task worktree and filter by status right from the command center, no more manually branching.Source →

Codex rust-v0.156.0 Codex stops losing progress when a turn is interrupted

Codex now preserves streamed answers and plans even when a turn fails, gets interrupted, or a subagent finishes. Resuming a session restores Plan mode, and editing an earlier prompt no longer loses thread identity or settings.

What changes If your connection drops or you accidentally hit stop mid-task, you won't have to rebuild the plan and answer from scratch.Source →

Claude app Claude app adds Opus 5.5

The Claude app added Claude Opus 5.5, the first model in the new Claude 5.5 family. It matches Fable 5.1-level performance on most tasks while costing 40% less to run than Opus 5.

What changes Tasks that called for Opus, like summarizing long documents or complex coding, are now cheaper to run than before.Source →

Claude Code 2.1.277 Claude Code's /ultrareview now refuses to run without a base branch

In non-interactive sessions, the /ultrareview command now refuses to run when the repository has no base branch or shared commit history.

What changes If you run /ultrareview inside a CI pipeline where nobody's watching, a repo with nothing to diff against will now fail loudly instead of silently reviewing against nothing.Source →

Claude Code 2.1.277 Claude Code's VS Code extension gets a Sign Out option

The VS Code extension's panel menu now has a Sign Out row, and typing /logout does the same thing.

What changes If you switch between accounts, you can sign out right from the panel instead of restarting the extension or dropping to a terminal.Source →

Claude Code 2.1.277 Claude Code's VS Code extension now tracks background shells in one map

The VS Code extension's agent map now lists background shells and other running tasks, each with a Stop button, and typing /tasks opens this list directly.

What changes When you're running several long sessions at once, you can see and stop background work from one screen instead of hunting across windows for what's still running.Source →

ChatGPT ChatGPT gets a single Privacy Center for all your settings

OpenAI is rolling out a Privacy Center to ChatGPT Free, Go, Plus, and Pro users that gathers chat privacy, memory, personalization, data use, connected apps, and account security settings in one place, reachable from the account menu on web or Settings on mobile.

What changes You no longer have to hunt through separate menus to find memory or connected-app permissions. If you use ChatGPT through a work account and need to check what data is being stored, this is the one screen to check first.Source →

ChatGPT ChatGPT can now show your credit score alongside financial advice

ChatGPT can now connect to your Experian credit report and VantageScore 3.0 score. The score and report refresh monthly, and you get alerts for changes like a new inquiry or account.

What changes You can connect just the credit-score piece without linking your bank accounts, so when you ask ChatGPT about a big purchase, it can factor in where your credit actually stands.Source →

Claude Code 2.1.277 Claude Code's /ultrareview Now Tells You What to Do Next When There's Nothing to Review

When /ultrareview finds nothing to review, it now explains which case you're in and suggests a command to review your latest commit instead. A new repository's first commit gets a full review.

What changes Instead of a dead-end 'nothing to review' message right after starting a branch, you get a concrete next command, cutting the time spent figuring out what to try next.Source →

Claude Code 2.1.277 Claude Code's Artifact Permission Prompts Now Show Real Titles

Artifact tool permission prompts now use shorter sentences, show pages and artifacts by their title or file name instead of a path, and list links after the main text.

What changes When Claude Code asks permission to open or publish an artifact, you can now tell what it is from its actual title instead of guessing from a file path.Source →

Claude Code 2.1.277 Fable Now Always Shows Up in Claude Code's Model List

Fable now always appears in the /model list when using the Anthropic API, and is only greyed out if your organization's settings disable it.

What changes Unless your organization has blocked access, you can now find and pick Fable directly from /model. Previously it could look entirely absent instead of just restricted.Source →

Claude Code 2.1.277 Claude Code Now Labels Subagent Output So It Can't Pass as Your Own Instructions

Subagent results now reach the main agent under a header marking them as subagent output, indented, so text inside a subagent's result can't be mistaken for the session's own instructions.

What changes If you run multi-agent workflows, it's now clearer on screen which text came from a subagent's output versus your own original instructions.Source →

Notion Notion 3.7 brings shared 'agent skills' for your whole team

Notion 3.7 adds agent skills that an entire team can share and use together.

What changes If your team keeps repeating the same workflow, you can now turn it into an agent skill and share it with everyone.Source →

Notion Notion 3.6 adds external agent connections and HTML blocks

Notion 3.6 lets you connect external AI agents into your workspace and adds a block for embedding raw HTML directly in a page.

What changes You can now call an AI agent you already use elsewhere directly from inside a Notion page, instead of switching tools.Source →

Notion Notion settings now let you connect Mail & Calendar directly

You can now connect your Mail and Calendar accounts straight from Notion's settings screen.

What changes You can check your schedule and mail without switching away from Notion.Source →

Notion Notion lets you customize view tabs in databases

You can now reorder and rename the view tabs at the top of a database however you like.

What changes Pin the views you check most often at the front, and you'll spend less time hunting for the right one.Source →

Claude Code 2.1.278 Claude Code moves auto mode's classifier server-side to cut billing overhead

For Claude API and Enterprise users, and on Bedrock, Vertex, Foundry, and gateways, auto mode now defaults to a server-side classifier that doesn't charge for the classification step itself, warns if it falls back to a billed path, and can be turned off with CLAUDE_CODE_AUTO_MODE_SERVER=0.

What changes If you leave auto mode on, your classifier costs drop with no setup needed, and you can revert with a single environment variable if you prefer the old behavior.Source →

Claude Code 2.1.278 Claude Code's /status now shows whether auto mode is running server-side

The /status screen now has a row showing whether this session's auto mode classifier is running on the server.

What changes You can confirm the cheaper server-side classifier is actually active just by checking /status.Source →

Claude Code 2.1.277 Claude Code's gateway loopback errors now name the exact fix

The Claude apps gateway's loopback error messages now name the CLAUDE_GATEWAY_ALLOW_LOOPBACK environment variable directly.

What changes When a loopback error shows up, you no longer have to search for the fix - the message names the exact variable to set.Source →

Claude Code 2.1.277 Claude Code strips invisible Unicode characters and shows the cleaned prompt before sending

Claude Code now strips invisible Unicode formatting and tag characters hidden in a prompt, and shows you the cleaned version before it's sent.

What changes If pasted text carries hidden characters you can't see, you can now review the cleaned-up version before it goes out.Source →

Claude Code 2.1.277 Claude Code now opens claude.ai artifact links with the Artifact tool directly

When you paste a claude.ai artifact link, Claude Code now reads it with the dedicated Artifact tool instead of WebFetch, whenever that tool is available.

What changes Sharing an artifact link is less likely to come back garbled, so you won't have to re-explain what's in it.Source →

Claude Code 2.1.277 Claude Code's dangerous-rm prompt now suggests a concrete fix

When Claude Code blocks a dangerous rm command, it now names the exact command that triggered it and suggests a ${VAR:?} guard so headless runs can recover on their own.

What changes If a headless run gets blocked on an rm command, adding the suggested guard lets the next run pass without you stepping in.Source →

Codex rust-v0.155.1 Codex Reverts a Default That Was Causing Request Rejections

Codex reverted reasoning-summary display in local TUI sessions back to off by default, fixing an issue where requests were being rejected by providers that don't support the feature. Anyone who explicitly turned it on keeps their setting.

What changes If you were hitting mysterious request rejections while using a model or provider that doesn't support reasoning summaries, this update alone should fix it.Source →

Codex rust-v0.155.0 Codex Adds Experimental Voice Conversations

An experimental /voice mode was added, letting supported builds hold spoken conversations with Codex complete with live transcripts and microphone controls, once enabled under /experimental.

What changes Worth trying if you'd rather talk through a task than type it, but it's experimental and has to be turned on manually first.Source →

Codex rust-v0.155.0 Codex's Status Row Now Shows Live Reasoning and Completion Time

The terminal UI's status row now streams live reasoning summaries while Codex works, and shows a completion timestamp once a turn finishes successfully.

What changes If you step away during a long task, you can tell what Codex was doing and when it finished just by glancing at the screen when you get back.Source →

Codex rust-v0.155.0 Codex's Agent Overview Adds Hide, Archive, and Delete

The agents overview screen, used when running several tasks at once, now supports hiding, archiving, and deleting tasks, and shows who owns each worktree. Deleting a clean managed worktree now asks for confirmation first.

What changes If you run several sessions at once, it's now easier to clear finished tasks out of the way and see only what's still in progress.Source →

Codex rust-v0.155.0 Codex Adds Touch ID Verification for MCP Requests on Mac

On supported Macs, local TUI sessions can now require Touch ID verification before sending an MCP request.

What changes For sensitive MCP tools, you can confirm with a fingerprint instead of a password every time, cutting both risk and friction.Source →

Codex rust-v0.155.0 Codex's Daemon Gets Schedulable Updates and Survives Restarts

You can now configure when Codex's background daemon updates itself, and the app-server daemon is now covered by updates too. Saved threads and active goals now survive a daemon restart.

What changes An update is less likely to cut off a thread you were mid-task on, so you shouldn't have to re-explain a long-running task from scratch.Source →

Codex rust-v0.155.0 Codex Lets You Manage Amazon Bedrock Credentials via a Command

When using Amazon Bedrock, you can now configure a command that fetches AWS credentials directly. Credentials are cached, refreshed automatically on expiry, and Codex attempts recovery if authentication drops.

What changes Teams with their own credential system on Bedrock can keep using Codex without manually refreshing credentials.Source →

Claude Code 2.1.277 Claude Code Now Falls Back to AGENTS.md

If a project has no CLAUDE.md, Claude Code now reads AGENTS.md instead, and you can change which file it uses under Project Instructions in /config. This isn't yet available on Bedrock, Vertex, or Foundry.

What changes If your project already uses the AGENTS.md standard for other AI coding tools, you no longer need a separate instructions file just for Claude Code.Source →

Claude Code 2.1.277 Claude Apps Gateway Adds an Egress-Boundary Option

A new CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1 option was added for gateways whose only egress path is a single forward proxy; with it on, every outbound request hands the proxy the hostname instead of resolving it locally.

What changes Teams that route all outbound traffic through a single proxy can now make Claude Code follow that same policy.Source →

Claude Code 2.1.277 Claude Apps Gateway Upstreams Can Now Send Static Headers

Gateway upstream configs now support an optional headers map, letting you send static header values to a proxy you run in front of a provider.

What changes If your internal proxy requires a custom header for authentication, that can now be handled entirely through gateway configuration.Source →

Claude Code 2.1.277 Background Task Completion Now Shows Even with a Panel Open

If a background task finishes while a panel like /tasks is open, Claude Code now shows a line saying its update is waiting.

What changes You can run several background tasks while looking at another panel and still notice the moment one finishes.Source →

Claude Code 2.1.277 Faster First Turn for SDK and Headless Use

When using Claude Code via the SDK or headless (-p) mode, the first turn no longer waits on the per-directory CLAUDE.md lookup before starting.

What changes If you call Claude Code repeatedly from an automated pipeline or script, the wait before the first response comes back gets shorter.Source →

Claude Code 2.1.277 /plugin's Install List Now Shows Which Plugin Owns Each MCP Server

In /plugin's installed list, an MCP server that's listed separately from its plugin now shows which plugin it belongs to.

What changes If you've installed several plugins and accumulated MCP servers, it's now easy to tell which one came from where.Source →

Claude Code 2.1.277 Reinstalling a Plugin Now Points You to the Update Command

Running claude plugin install on a plugin that's already installed now tells you if the marketplace has a newer version, and names the claude plugin update command to use.

What changes No more reinstalling out of uncertainty and getting no feedback; you can tell right away whether an update is actually needed.Source →

Codex rust-v0.155.1 Codex's terminal UI reverts its reasoning-summary default to off

Codex's terminal UI reverted the default for showing reasoning summaries back to "none" (off).

What changes If your terminal suddenly stopped or started showing reasoning summaries without you changing anything, this default flip is why. Adjust the setting again if you want a different display mode.Source →

Claude Code 2.1.275 Claude Code Now Shows Which Account You're Signing Into via Gateway

Claude Code now shows the signed-in account the gateway reports before saving the credential, and /status displays it.

What changes If you switch between accounts through a company gateway, you can catch a wrong-account sign-in right at login instead of discovering it mid-session.Source →

Claude Code 2.1.275 Claude Code Adds a Key to Send All Queued Messages at Once

Claude Code added a send-now key (ctrl+enter, or ctrl+x ctrl+s) that interrupts the current turn and sends all queued messages at once; sent and queued messages show in gray until the model receives them.

What changes If you tend to type up several messages while waiting for a long task to finish, you can now send them all at once instead of waiting turn by turn.Source →

Claude Code 2.1.275 Claude Code Now Syncs Your claude.ai Skills and Plugins to the Terminal

Claude Code now syncs the skills and plugins enabled on your claude.ai account to terminal sessions signed in with that account; opt out with syncClaudeAiSkills: false or syncClaudeAiPlugins: false.

What changes If you've ever enabled a skill on the web only to find it missing in the terminal and had to set it up again, that extra step is gone.Source →

Claude Code 2.1.275 Claude Code's Plugin Install Command Can Now Add the Marketplace Itself

Claude Code added /plugin install <plugin> --marketplace <source>, which offers to add the marketplace before installing the plugin.

What changes If you used to look up a separate command just to register a new marketplace, installing straight from it now takes one command.Source →

Claude Code 2.1.275 Claude Code's /desktop Command Now Explains Why It Failed

Claude Code improved the /desktop error when Claude Desktop does not open: it now says why and what to do next.

What changes When /desktop used to just hang with no explanation, you had to guess and restart; now it tells you what actually went wrong.Source →

Codex rust-v0.155.0 · Bug Fixes Codex Now Clears Old Session Data When You Switch Accounts

Switching accounts in Codex now invalidates the previous account's remote-control sessions, cached WebSocket state, and model catalog.

What changes If you switch between personal and work accounts in Codex, you'll see fewer cases where stale state from the previous account causes unexpected behavior.Source →

Codex rust-v0.155.0 Codex now keeps your approved prompt even if compaction fails

If context compaction fails before a turn starts, your already-accepted prompt is now saved instead of being lost.

What changes If compaction used to fail mid-session and wipe out the prompt you just wrote, you won't have to retype it anymore.Source →

Codex rust-v0.155.0 Codex now explains why an MCP connection broke and how to reconnect

When an MCP server's OAuth credentials expire and token refresh also fails, Codex now reports that accurately and walks you through reconnecting.

What changes When an MCP tool silently stops working, you won't have to guess why anymore.Source →

Codex rust-v0.155.0 Codex closes a WSL sandbox escape and a credential-exposure gap

Codex blocked a way for processes to escape restricted WSL sandboxes into Windows, and hardened brokered shell snapshots against exposing credentials.

What changes If you run Codex sandboxed inside WSL on Windows, that isolation just got tighter with no action needed on your part.Source →

ChatGPT ChatGPT Now Works Directly Inside Word

ChatGPT is now available inside Microsoft Word as a sidebar, letting you draft from notes, summarize a document, revise selected text, or adjust heading formatting. Word joins Excel and PowerPoint through the same add-in.

What changes When you're drafting a report or proposal in Word, you can revise it right from the sidebar instead of copying text back and forth with a separate ChatGPT window.Source →

ChatGPT ChatGPT Can Now Use Personal and Work Accounts in the Same Chat

ChatGPT now lets you connect multiple accounts to plugins, expanding beyond Gmail, Google Calendar, and Google Contacts. You can bring both personal and work accounts into the same conversation.

What changes If you've been checking work and personal calendars separately, you can now ask ChatGPT about both accounts in the same conversation.Source →

Claude Code 2.1.273 Claude Code now explains exactly why a GitHub connection failed

When a cloud session's GitHub connection is blocked by an IP allow list, a suspended app installation, or SAML single sign-on, Claude Code now shows the specific cause. The same applies when gh pr view fails or webhook setup fails inside /autofix-pr, and when a GitHub token is refused inside /web-setup.

What changes The error used to say something vague like "check your installation," leaving you to track down the cause yourself. Now the reason shows up directly, so you can tell your IT admin exactly what needs to be unblocked.Source →

Claude Code 2.1.273 Claude Code now names the fix for SSL and proxy connection errors

When an in-session SSL certificate or proxy connection error is caused by an untrusted corporate proxy or CA, Claude Code now names the error code and what to fix, such as setting NODE_EXTRA_CA_CERTS.

What changes An unexplained SSL error on a corporate network used to eat up time just tracking down the cause. Now the fix, like which environment variable to set, shows up directly, so you can often resolve it yourself without waiting on IT.Source →

Claude app Claude app brings Cowork into every conversation, no mode to pick

Everything Claude Cowork does is now available from any conversation, without picking a mode first. You can move between a quick question and a bigger task, like a report, spreadsheet, or presentation, in the same chat, and your chats, Cowork tasks, projects, connectors, and skills all carry over. It's rolling out gradually to Pro and Max plans on web, desktop, and mobile.

What changes If a side question comes up while you're building a report, you don't need to open a new chat, you can just keep asking in the same conversation.Source →

Claude app Claude can now create designs, slides, and docs mid-conversation

You can now ask for a design, deck, or document in any conversation with Claude, including in Claude Code and the Artifacts tab, and edit it directly on the canvas. Claude Design, Claude Slides, and Claude Docs are available on every plan including Free; on Enterprise plans they're in beta and off by default until an admin turns them on.

What changes When you need a slide deck or planning doc, you don't have to switch tools, you can make and edit it right inside the Claude conversation you're already in.Source →

Notion Notion rollups can now display as currency or percent

Notion database rollups, which aggregate values from another table, now support new display formats including currency and percent.

What changes If you track revenue or progress with rollups, you can now read the numbers directly as currency or percent in the table without reformatting them elsewhere.Source →

Notion Notion lets you search pages right from the breadcrumb

Clicking a page's breadcrumb trail now lets you search for and jump to another page right from that spot.

What changes Instead of clicking up through parent folders and hunting in the sidebar, you can jump straight to another page from wherever you already are.Source →

Notion Notion AI now offers Claude Opus 4.7

Notion AI's model picker now includes Anthropic's Claude Opus 4.7.

What changes For more demanding writing or research tasks, you can now pick a stronger model right inside Notion instead of switching to another tool.Source →

Notion Notion Agent can now schedule meetings for you

Notion Agent can now schedule meetings directly within a conversation.

What changes On a day packed with meetings, you can set up the next one by chatting inside Notion instead of switching over to a calendar app.Source →

Claude Code 2.1.273 Claude Code now flags it when an MCP server disconnects

When an MCP server disconnects mid-session and automatic reconnection gives up, Claude Code now shows a notification pointing you to the `/mcp` command.

What changes When a tool connected through MCP suddenly stops responding, you can jump straight to the reconnect command instead of wondering what went wrong.Source →

Claude Code 2.1.273 Sessions started via remote control can now be forked

You can now fork a session started with `claude --remote-control` or `/remote-control`; the fork runs as a background session on your own computer.

What changes Even if you started a session remotely from the Claude app on your phone, you can branch off work from it and keep it running in the background on your own computer.Source →

Claude Code 2.1.273 Claude Code stays responsive in long sessions

Hook progress and sub-agent activity no longer re-process the entire conversation on every update, improving responsiveness in long sessions.

What changes If long sessions running multiple sub-agents used to stutter, that lag should now be noticeably reduced.Source →

Claude Code 2.1.273 Artifact database writes can now remove a single field

An artifact database update can now remove a single field instead of rewriting the entire document.

What changes If you're building an artifact that stores data, like a checklist or notes app, deleting a single entry is now a lighter, safer operation.Source →

Claude Code 2.1.273 A dropped artifact publish now retries safely, no duplicates

If the connection drops after a publish reaches claude.ai, it used to risk failing or creating a duplicate version. It's now resent safely instead.

What changes On a shaky connection, publishing a dashboard or page no longer risks ending up with duplicate versions after you retry.Source →

Claude app Salesforce data now works directly inside Claude

Claude has added a Salesforce plugin in beta that pulls a seller's accounts, opportunities, and pipeline directly into the chat. With 37 pre-built sales skills, you can prep a call, review a deal, build a pipeline dashboard, or send a forecast without leaving Claude.

What changes If your sales team uses Salesforce, you can skip opening a separate tab before a meeting and just ask Claude about an account to get the latest status.Source →

Claude Code 2.1.271 Fast mode now works in Claude Code Remote sessions

Claude Code Remote sessions, whether on cloud or self-hosted runners, can now use fast mode. It follows the host's fast-mode setting or can be turned on with /fast typed directly in the session, wherever the organization allows it.

What changes When running long tasks on a remote runner, typing /fast now gets you faster output there too.Source →

Claude Code 2.1.271 The /config panel now supports mouse clicks and scrolling

Mouse support has been added to the /config panel in fullscreen mode. The wheel scrolls the settings list, clicking a setting's value changes it, and the row under the pointer is highlighted.

What changes Instead of navigating settings one arrow-key press at a time, you can now click straight to the setting you want to change.Source →

Claude Code 2.1.271 Sandboxed auto mode can now scope allowed domains per command

Sandboxed auto mode now supports per-command allowed_domains for Bash, PowerShell, and Monitor. The hosts a command actually needs are reviewed and opened for it alone; requests to any other host are refused.

What changes When running network-touching commands in auto mode, you can now scope approval to just the domains that command actually needs, instead of opening access more broadly.Source →

Claude Code 2.1.271 Subagents can now skip CLAUDE.md files entirely

A new omitClaudeMd option in agent frontmatter and --agents JSON lets custom and plugin subagents run without loading user, project, or local CLAUDE.md files. Managed policy files still load regardless of this setting.

What changes When building a standalone subagent where the project's CLAUDE.md instructions would only get in the way, this option lets it run on a clean context without them.Source →

Claude Code 2.1.271 Plugin installs can now be approved by exact command hash instead of -y

claude plugin install and claude plugin update now support --accept-command <sha256>, letting you approve the exact command a previous --json run displayed, by its hash, instead of blanket-approving with -y.

What changes Instead of blanket-approving everything with -y in install scripts, you can now require an exact match to a command you've already reviewed, which is safer in automated pipelines.Source →

Claude Code 2.1.271 A new spinner tip nudges gateway users toward the Claude desktop app

Users running Claude Code through Bedrock, Vertex AI, Foundry, or an LLM gateway now see a spinner tip pointing them to the Claude desktop app. For claude.ai desktop app users, the tip now suggests typing /desktop, which offers to download the app.

What changes If you've only been using the CLI through a gateway setup, this tip may be the first sign that a desktop app option exists.Source →

ChatGPT ChatGPT stops auto-switching to its thinking mode

For ChatGPT Plus and Pro users worldwide, automatic switching from Instant to Thinking mode is being retired. You'll now need to manually pick Thinking mode from the model picker, and the "Higher intelligence" setting is being removed too.

What changes If you were relying on ChatGPT to notice a hard question and think it through automatically, you'll now need to switch to Thinking mode yourself from the model picker whenever an answer seems shallow.Source →

ChatGPT ChatGPT no longer asks permission every time to use your Health data

ChatGPT now defaults new Health plugin connections to "Allow low-risk actions," so once you connect your Health data, ChatGPT can use it without asking each time, unless you're asking for something sensitive.

What changes If you've already connected Health data, ChatGPT may start using it without asking starting today. If that feels like too much, you can lower the permission yourself in Settings → Plugins → Health.Source →

Claude Code 2.1.269 Claude Code fixed the slowdown in long sessions

Claude Code no longer reprocesses the entire conversation history just to rebuild collapsed tool-use summaries, making long sessions noticeably more responsive.

What changes If you run long sessions for hours, the lag and input delay you used to feel should ease up. It'll be most noticeable when you keep one window open across many tasks in a row.Source →

Claude Code 2.1.269 /ultrareview --post now posts straight to the PR instead of spinning up a second session

The /ultrareview --post command now posts the PR comment directly as soon as findings arrive and prints the comment link, instead of spinning up a second cloud session just to post it.

What changes You get the PR comment faster and without the cost of a second session. As soon as the review finishes, you get a link to the comment right away.Source →

Claude Code 2.1.269 Reading artifact database results into the scratchpad no longer asks for folder approval

When artifact database reads save their results into the session scratchpad, Claude Code no longer pauses to ask for working-folder approval before writing the file.

What changes The approval popup that used to interrupt every artifact data pull is gone, so working with large documents or many files no longer breaks your flow.Source →

Claude Code 2.1.269 Skills synced from claude.ai in cloud sessions now get an anthropic-skills: prefix

Skills synced from claude.ai in cloud sessions are now named anthropic-skills:<name>, matching Claude Desktop. The bare name still works as long as nothing else uses it.

What changes The name you're used to typing may still work, but if more than one skill shares that name, you'll now need the anthropic-skills: prefix to call the right one without ambiguity.Source →

Claude Code 2.1.269 Claude Code now hints at a new /focus command

A new spinner tip now points to the /focus command, which switches to a view showing just your prompt, a one-line summary of the work, and the response.

What changes When the screen gets cluttered with tool calls and logs, /focus lets you see just what's actually happening right now.Source →

Claude Code 2.1.269 Claude Code's Workflow tool lets you raise the concurrent-agent limit yourself

The new CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS environment variable lets you set how many agents a single Workflow run can execute at once, from 1 to 256.

What changes For large fan-out tasks that need many agents running in parallel, you're no longer stuck waiting behind the default concurrency cap.Source →

Claude Code 2.1.269 Claude Code's /diff panel now opens fully rendered, no loading screen

The /diff panel now opens fully rendered in one step instead of flashing a loading state first.

What changes If you check diffs often, the panel now feels smoother since that brief loading flash is gone.Source →

Claude Code 2.1.269 Claude Code improves prompt-suggestion filtering for Korean, Japanese and Chinese

For Korean, Japanese and Chinese input, Claude Code now keeps mixed-script and single-word prompt suggestions while dropping meta or evaluative ones, matching how it already handled English.

What changes Korean-language users should see more useful autocomplete suggestions in Claude Code, an area that had lagged behind the English experience.Source →

Claude Code 2.1.269 Claude Code fixes broken shortcuts over SSH and in some terminals

Shift+Enter and Ctrl+Shift shortcuts now work in terminals that support the kitty keyboard protocol, such as foot and Alacritty 0.16+, including over SSH and in less common terminals.

What changes If you use Claude Code over SSH, shortcuts that didn't work before should now behave normally.Source →

Claude Code 2.1.269 Claude Code Can Now Score a Plugin's Own Eval Suite

The new 'claude plugin eval' command runs a plugin's own eval suite against Claude Code and returns scored, reproducible results as a JSON and HTML report.

What changes If you build your own plugin, you can check whether each update still works from one report instead of testing it by hand every time.Source →

Claude Code 2.1.269 Claude Code Lets You Switch Output Styles Even in Remote or Cloud Sessions

The new '/output-style [name]' command lists and switches output styles, and now also works over Remote Control and in cloud or other headless sessions.

What changes Even when you're running a remote session from your phone, you can switch output styles the same way you would at your computer.Source →

Claude Code 2.1.269 Claude Code Shows a Diff When a Bash Command Edits Files

When the Bash tool edits files directly, Claude Code now includes a diff of what changed in the tool result (via the 'bashEditDiffEnabled' setting).

What changes When a single command edits several files, you can see exactly what changed without opening each file yourself.Source →

Claude Code 2.1.269 Windows: Background PowerShell Commands No Longer Die When Claude Code Exits

Fixed a Windows bug where PowerShell commands sent to run in the background would stop as soon as Claude Code exited.

What changes On Windows, a long build or server you started in the background keeps running even after you close Claude Code.Source →

ChatGPT ChatGPT is retiring custom GPTs in favor of plugins

OpenAI is planning to retire custom GPTs across all ChatGPT plans and is building a migration path to "plugins," which combine reusable instructions with connected apps. Existing GPTs will keep working until each one's retirement date, subject to permissions.

What changes If you rely on a custom GPT at work, review now which ones you use, who created each one, and who needs access, so you're not scrambling when the migration date hits.Source →

Notion Notion Lets You Choose Which AI Model Your Agents Use

Notion now lets you specify which AI model each of your agents uses.

What changes You can assign a lighter model to tasks where speed matters and a stronger one to tasks that need careful judgment.Source →

Notion Notion Lets You Share Context With Custom Agents From the Share Menu

You can now hand a page's context straight to a Custom Agent from that page's Share menu.

What changes You no longer have to open the agent separately and re-explain the document. You can just keep asking questions right from the page you're already viewing.Source →

Notion Notion Workers Can Now Be Shared Across Your Team

You can now share a Notion Worker, an automated workflow you've built, with your teammates so they can use it too.

What changes One person's automated workflow no longer has to be rebuilt by everyone else on the team. They can just reuse it.Source →

Notion Notion Agents Gets Its Own iOS App

Notion released a standalone iOS app just for using its agents.

What changes You can assign work to an agent and check its results on the go, without opening the main Notion app.Source →

Claude Code 2.1.268 Claude Code Adds --json to Plugin Commands

The claude plugin install, uninstall, update, enable, and disable commands now support --json, and each row of claude plugin list --json carries error and note details.

What changes When installing or managing plugins from a script, you can now read success or failure as structured data instead of parsing human-readable text.Source →

Claude Code 2.1.268 Claude Code Adds Browser-Tab Icons to Published Artifacts

When you publish an artifact, Claude now picks a matching icon to show in the browser tab.

What changes With several artifact tabs open at once, you can tell them apart by icon alone instead of reading every title.Source →

Claude Code 2.1.268 Claude Code's Fullscreen Mode Redraws Faster While You Type

In fullscreen mode, adding or removing a prompt line with Shift+Enter now repaints as fast as typing a single character, instead of re-rendering the whole visible transcript.

What changes Typing a multi-line prompt in fullscreen during a long session no longer causes the screen to stutter.Source →

Claude app Claude Now Generates Reports on How Your Team Actually Uses It

Claude launched Smart Reports in beta, which analyze how a team uses Claude and report on what work is getting done, what it costs, where sessions hit friction, and which repeated patterns are worth packaging as shared skills. It's available on Claude Enterprise plans.

What changes If you're the one managing how your team uses Claude, you can now see where people are getting stuck from an actual report instead of guessing.Source →

Gemini The Gemini App Comes to Windows

Google released a Gemini desktop app for Windows, built to run alongside the other programs you use every day.

What changes You can now pin Gemini to your desktop instead of switching browser tabs, handy for asking questions right next to the document you're working on.Source →

Codex python-v0.154.0 Codex Adds Max and Ultra Reasoning-Effort Options

Codex added new 'max' and 'ultra' reasoning-effort levels on top of its existing options.

What changes For a hard bug that needs deeper reasoning, you can now dial the effort up a notch further than before. Skip it for quick, simple tasks, since it costs more time and money.Source →

ChatGPT ChatGPT Work and Codex Get a Data Plugin for Business Questions

ChatGPT Work and Codex now have a Data plugin that analyzes connected business data, letting you ask a question and get an investigation, dashboard, or report back.

What changes Install Data from the plugin directory and mention @Data in a conversation to use it. Instead of opening a fresh spreadsheet every time, you can just ask about your metrics in chat.Source →

ChatGPT ChatGPT Library Now Connects Box, Dropbox, and SharePoint

Box, Dropbox, and SharePoint now join Google Drive in ChatGPT Library, letting you browse and search connected files and folders directly in ChatGPT.

What changes You can pull a file straight into a conversation via Add from Library or an @mention, without re-uploading it. If your team's documents are spread across several clouds, this saves the step of copying them over each time.Source →

Claude Code 2.1.267 New maxEffortLevel Setting Caps Reasoning Effort Across Providers

Setting maxEffortLevel, either top-level or per model under modelSettings, caps the reasoning effort level across every provider including Bedrock, Vertex, and Foundry; users can still choose a lower level.

What changes If you need to manage cost or response speed across a team, you can set one organization-wide cap instead of asking everyone to configure it session by session.Source →

Claude Code 2.1.267 Claude Now Suggests /copy When Clipboard Commands Fail in the Sandbox

When clipboard commands like pbcopy fail inside the sandbox, Claude now suggests using /copy instead.

What changes When copying suddenly stops working in the sandbox, you no longer have to guess why; Claude points you straight to /copy instead.Source →

Claude Code 2.1.267 Faster --resume for Sessions With Many Bash Tool Calls

Resuming a session with many accumulated Bash tool calls now renders its first screen faster.

What changes The longer a session's Bash history, the less you now wait after --resume before you can pick up where you left off.Source →

Claude Code 2.1.267 Fixed Keystrokes Occasionally Lagging Behind the Spinner

Prompt input responsiveness was improved: keystrokes no longer occasionally wait a frame behind spinner or streaming repaints.

What changes Typing while Claude is streaming a long answer no longer occasionally looks laggy or glitchy on screen.Source →

Codex rust-v0.154.0 Codex now keeps your permission settings intact when you resume or fork a session

Resuming or forking a remote Codex session now keeps whatever permissions you had saved. Brand-new sessions and forks fall back to the server's default model unless you explicitly set one.

What changes If you used to have to re-approve permissions every time you resumed a session, that's no longer necessary since your settings now carry over automatically.Source →

Codex rust-v0.154.0 Codex won't lose your draft anymore when you resume a conversation open in another app

If you resume a conversation that's already open in another app, Codex now shows a read-only transcript with a retry option, while keeping whatever draft you were writing intact.

What changes If you keep the same conversation open on two devices at once, you no longer risk losing what you were typing. Just hit retry instead.Source →

Codex rust-v0.154.0 · Bug Fixes Codex now picks up plugin updates in sessions you already have open

Existing Codex sessions now pick up newly installed plugin tools and refresh skills and hooks right after an external plugin is upgraded or rolled back.

What changes Previously you had to restart the session after updating a plugin. Now you can keep working in the same window and the updated tool is already there.Source →

Codex rust-v0.154.0 · Bug Fixes Codex blocks untrusted startup helpers and a macOS terminal input-injection hole

Codex now avoids running workspace-controlled helper programs before trust is established, and its macOS sandbox blocks terminal input injection.

What changes Opening an unfamiliar repository is safer, since workspace-controlled code can no longer run before you've confirmed you trust it. It's one more safeguard when working in an untrusted codebase.Source →

Codex rust-v0.154.0 · New Features Codex's Vim Mode Gets an Undoable Replace Command

Codex CLI's vim editing mode now supports the R (replace) command with undo and dot-repeat, and Escape key handling in legacy terminals is more reliable.

What changes If you're editing code with vim keybindings and accidentally overwrite text, you can now undo it in one step instead of retyping.Source →

Codex rust-v0.154.0 · New Features Copying a Codex Response Now Keeps Its Formatting

Copying a Codex response now preserves formatting when pasted into rich-text apps, and the /copy command can copy just the status output or specific session fields.

What changes When you paste a code explanation or table Codex generated into Slack or a doc, you won't need to fix broken formatting afterward.Source →

Codex rust-v0.154.0 · New Features Codex on Windows Can Now Share One Background Server Across Sessions

Codex sessions on Windows can now share a single background server, with commands to manage its lifecycle and handle updates automatically.

What changes Running Codex in multiple windows no longer spins up duplicate servers, so it uses less memory on Windows and starts up faster.Source →

ChatGPT Deep Research Arrives in ChatGPT Work and Codex

Deep Research, which investigates a complex question across the web, your files, and connected apps, then turns the findings into a cited document, is now available in ChatGPT Work and Codex. You can redirect it mid-research.

What changes Type @Deep Research or just ask for it, and you can get results back as a document, presentation, spreadsheet, or site. It's worth trying for tasks like competitor research or market reports that need a wide sweep of sources.Source →

Codex rust-v0.154.0 GPT-6 Astra Is Now a Model Choice in Codex

GPT-6 Astra is now available in Codex's model picker and in the Amazon Bedrock catalog.

What changes For harder tasks, you can just switch to Astra from the model picker without any extra setup.Source →

Codex rust-v0.154.0 New --worktree Flag Gives Each Session Its Own Checkout

The --worktree flag or /worktree command creates an isolated checkout for a new or forked session, which you can later browse and resume.

What changes When you're running several tasks at once, you no longer have to juggle branches in one folder; each session gets its own fully separate directory.Source →

Codex rust-v0.154.0 Answer Codex's Mid-Task Questions Without Losing Your Draft

When Codex asks a question mid-task, you can answer inline with a suggested choice or custom text, without losing whatever you were drafting in the main input.

What changes When Codex asks which way to go, you can just answer on the spot and keep going, no need to clear or save what you were typing.Source →

Codex rust-v0.154.0 MCP Connections Now Prompt Re-Login When Token Refresh Fails

MCP connections now coordinate OAuth token refreshes and surface a login prompt when a refresh fails, instead of silently retrying a rejected tool call.

What changes Instead of a connected MCP tool silently failing for no visible reason, you now get a clear prompt to log back in and keep going.Source →

ChatGPT ChatGPT Voice Can Now Answer With GPT-5.6 or GPT-6 Astra

ChatGPT Voice can now use GPT-5.6 or GPT-6 Astra when it needs to search or reason through harder questions, with the same model and reasoning-effort controls as text chat. Plan-based usage limits were also simplified.

What changes You can now expect voice answers to complex questions to be as capable as text chat. Usage time still depends on your plan, so it's worth checking your limit in settings.Source →

ChatGPT Share Files and Folders From Your ChatGPT Library

You can now share files and folders from your ChatGPT Library, choose who can access them, and work with shared content directly inside conversations.

What changes Instead of re-attaching files for a colleague every time, you can share a Library folder once and keep referencing it across future conversations.Source →

Codex rust-v0.153.4 Codex adds GPT-6 Astra to its model picker and Amazon Bedrock

Codex added OpenAI's newer GPT-6 Astra model to its bundled model picker and to Amazon Bedrock's catalog, along with refined in-app guidance for Astra-specific features like asynchronous questions.

What changes You can pick Astra straight from Codex's model picker without extra setup, and teams deploying through Bedrock can now use the same model.Source →

Notion Notion lets you ask your agent to suggest document edits

In Notion, you can now ask an AI agent to propose edits to a document instead of changing it directly, so a person reviews and decides whether to accept each suggestion.

What changes If you've hesitated to hand a document to an agent for fear it would rewrite things outright, you can now just review its suggested edits and approve them yourself.Source →

Notion Notion simplifies its AI model picker

Notion simplified the list of AI models you can choose from when using its AI features.

What changes If you were never sure which model to pick, the shorter list means one less decision before you can just get started.Source →

Notion Notion's AI Meeting Notes can now kick off Custom Agents

When Notion's AI Meeting Notes finishes, it can now automatically kick off a Custom Agent to handle follow-up work.

What changes The task of reading meeting notes and manually entering action items after every meeting can now be handed off to an agent you've already set up.Source →

Notion Notion agents get calendar tools

Notion agents now have tools to check and adjust calendar events directly.

What changes Instead of checking back and forth with a calendar to coordinate a schedule, you can let an agent handle it.Source →

Claude Code 2.1.265 Claude Code caps saved tool results at 1 GB

Claude Code now caps tool results saved to disk at 1 GB, and the in-conversation preview notes when a saved file was truncated.

What changes Running a task that produces huge output no longer risks suddenly filling up disk space, and you're less likely to miss that a result got truncated.Source →

Claude Code 2.1.265 Claude Code fixes Windows sandbox blocking all file access

Fixed a Windows bug where running inside an AppContainer or restricted-token sandbox made Read, Write, and Edit refuse every file.

What changes If you were running Claude Code in a Windows sandbox environment, the situation where it couldn't read or write any file at all is now fixed.Source →

Claude Code 2.1.265 Claude Code shows a list of matches for slash commands typed mid-prompt

Typing a slash command in the middle of a prompt now shows matching commands as a list instead of a single suggestion; Tab opens it even outside fullscreen, and a plugin skill can be found by its bare name.

What changes When you're mid-prompt and can't quite remember a command's name, you can just pick it from the list instead of retyping.Source →

ChatGPT ChatGPT images get templates and sketch-to-image

ChatGPT's image generation moved to version 2.5 with sharper output and more precise editing. Users can now start from a template, turn a hand-drawn sketch into an image on mobile, or comment directly on a generated image to request changes.

What changes When it's hard to describe an idea in words, starting from a rough sketch or a ready-made template gets you to the result you want faster.Source →

Notion Notion adds a high contrast mode

Notion added a high contrast display mode to its settings.

What changes If low contrast made Notion hard to read, you can now turn this on in settings for a clearer display without a browser extension.Source →

Claude Code 2.1.261 Claude Code now shows why your org policy failed to load, right in /status

Added an "Organization policy" line to /status and claude doctor that explains why your organization's policy failed to load, such as a proxy not passing the endpoint through.

What changes Instead of guessing why company policy isn't applying, running /status now tells you the actual cause, like a proxy misconfiguration, so you can hand IT the exact problem.Source →

Claude Code 2.1.261 You can now raise the character limit on long command output

Added bashOutputMaxChars and taskOutputMaxChars settings to raise how much command and background-task output Claude receives inline before it's saved to a file, up to 128K characters.

What changes If long test logs or build output used to get cut off mid-way, raising this setting lets Claude see the whole thing at once instead of missing context.Source →

Claude Code 2.1.261 Long subagent prompts can now be passed as a file

Added --append-subagent-system-prompt-file to read the subagent system prompt from a file, for prompts too large to pass on the command line.

What changes If you build custom subagents, you can now hand over a long instruction set as a single file instead of trimming it to dodge shell-escaping issues.Source →

Claude Code 2.1.261 /skill-doctor shows which unused skills are eating your context

Added /skill-doctor to show which loaded skills go unused and what they cost in context, so you can prune them.

What changes If a session feels sluggish with several skills loaded, /skill-doctor can point out the unused ones so turning them off frees up context.Source →

Claude Code 2.1.261 /model picker now shows model names instead of raw IDs

Improved the /model picker and the VS Code model pill to show a model's name instead of its raw Bedrock, Vertex AI, or LLM gateway ID when Claude Code recognizes it.

What changes If you use Claude through Bedrock or Vertex AI, you can tell which model is selected at a glance instead of decoding a long ID string.Source →

Claude Code 2.1.261 Faster startup on Google Vertex AI when credentials are pre-set

Improved startup on Google Vertex AI when GOOGLE_APPLICATION_CREDENTIALS is set: API client creation no longer re-runs Google Cloud project discovery or spawns extra gcloud processes.

What changes If Claude Code felt slow to start on Vertex AI, this fix skips the repeated gcloud calls on every run, cutting the wait.Source →

Claude Code 2.1.261 Long streaming responses no longer slow down the display

Improved streaming performance: already-rendered blocks are no longer re-checked by layout on each update.

What changes Long responses that used to make the terminal stutter should feel smoother, cutting the perceived wait during long outputs.Source →

Claude Code 2.1.261 The dangerous rm -rf warning now catches more disguised forms

Improved the dangerous-rm safety prompt to also catch rm -rf on positional parameters and inside double-quoted sh -c scripts.

What changes rm -rf commands hidden in variables or nested scripts now trigger a warning before running, adding one more check against accidental deletion.Source →

Codex rust-v0.153.3 Codex now offers GPT-6 Astra through Amazon Bedrock

GPT-6 Astra was added to Codex's Amazon Bedrock model picker and catalog, for Mantle and Runtime global/US routes.

What changes Teams required to route models only through Amazon Bedrock can now pick GPT-6 Astra directly in Codex, with no separate approval or API key changes needed.Source →

Claude Code 2.1.259 Claude Code narrows what allowedMcpServers governs (check on upgrade)

The allowedMcpServers setting now governs only MCP servers that users add themselves. A specific managed-mcp.json server that this allowlist used to filter out will now load automatically on upgrade; use deniedMcpServers to keep it off.

What changes If you were relying on allowedMcpServers to block a specific managed server, it may come back on after upgrading. Check your deniedMcpServers setting.Source →

Claude Code 2.1.257 Claude Code now prompts before its first read outside your working directory

In auto mode, Claude Code now shows a one-time prompt before its first file read outside your working directories, with a new permissions.blockReadsOutsideWorkingDirectories option to block such reads entirely.

What changes If you run auto mode across multiple projects, this setting lets you prevent it from reading files outside the current one by accident.Source →

Claude Code 2.1.259 Claude Code preserves nested background subagent results

Results from a background subagent launched inside another background subagent are now saved to the parent subagent's transcript, so resumed sessions keep them and shared transcripts show that delivery happened.

What changes If you nest background agents inside each other, you're less likely to lose track of a sub-task's result when you resume the session.Source →

Claude Code 2.1.259 Claude Code's VS Code extension adds status filters to the session list

The VS Code extension's session list sidebar now has an Active quick filter and a status filter menu (Needs input, Working, Completed).

What changes When you're running several sessions at once, filtering down to the ones waiting on input cuts the time spent hunting for what's happening where.Source →

ChatGPT ChatGPT adds Zendesk and OneNote plugins

Zendesk and OneNote plugins were added to the plugin directory in ChatGPT and Codex. Users connect their own accounts to pull in support tickets or notes.

What changes When preparing a support reply, you no longer need to switch between Zendesk and ChatGPT — you can pull up the ticket history and draft the reply inside ChatGPT itself.Source →

ChatGPT ChatGPT Sites can now be shared with people outside your workspace

Eligible Site owners can now share a live ChatGPT Site with named people outside their workspace by email, without making it public. Invited viewers can view but not edit or republish.

What changes You can open up a company ChatGPT Site to one outside partner or client without making it public, just by entering their email in the Share panel.Source →

Codex rust-v0.153.1 Codex lets you use GPT-6 Astra via API without switching your default model

Codex now lets you configure GPT-6 Astra through the API without changing your default model or having it show up in the model picker.

What changes You can try Astra's performance on a specific task without touching your team's default configuration.Source →

Codex rust-v0.153.0 Codex TUI now reconnects without losing your session

Codex TUI sessions now automatically reconnect after an external app-server connection drops, keeping drafts and transcripts intact, with uncertain or queued submissions paused for review.

What changes A brief network drop is less likely to wipe out a session you were in the middle of.Source →

Codex rust-v0.153.0 Codex Guardian review history now survives restarts and forks

Codex's Guardian review history now survives conversation compaction, restarts, and user-created forks, while still respecting rollback boundaries and keeping subagent history separate.

What changes Even after compacting or restarting a long session, you keep a record of what was approved, so you can trace back what happened later.Source →

ChatGPT ChatGPT begins rolling out GPT-6 Astra

ChatGPT is rolling out GPT-6 Astra, with improvements in coding, research, computer use, and complex multi-step work, starting with a limited set of organizations. It's not yet generally available, with broader rollout planned over the coming days.

What changes Astra can pause a conversation on its own and ask you to review it if it suspects it may have misread your instructions. If ChatGPT suddenly stops mid-task, that may be this safeguard rather than an error.Source →

Codex Codex Vim mode now has undo and redo

In Codex's Vim mode you can undo with u and redo with Ctrl+R. Pasted content and attachments are kept as part of the draft.

What changes If you wiped a long prompt by accident, you used to start over. Now one undo brings back the whole draft, including pasted code and files.Source →

Codex Install Codex plugins from remote marketplaces via the CLI

The plugin CLI can list, install, and remove plugins from remote marketplaces.

What changes Instead of downloading and placing plugins by hand, you can install and clean them up straight from a marketplace with one command.Source →

Codex You can turn off Codex's automatic recaps

Set tui.auto_recap = false to turn off automatic recaps. The manual /recap command still works when you want one.

What changes If the recaps that appear during long sessions got in your way, you can switch them off and call /recap only when you want it.Source →

Codex Codex history now shows full patches and commands

TUI history now shows complete patches, input sent to background terminals, and each completed command on its own.

What changes You can retrace exactly what Codex changed and which commands it ran, in full rather than a clipped summary. That helps when you check back or trace a problem.Source →

Codex Codex warns about your usage limit earlier

Plus and Team users get an earlier warning once less than half of their allowance is left in a roughly five-hour usage window.

What changes Instead of hitting the limit mid-task, you learn when about half is left and can plan when to run heavier work.Source →

Claude Code 2.1.259 Claude Code Adds an Org-Wide Managed Setting for MCP Servers

Claude Code added a managedMcpServers setting, letting organizations roll out HTTP/SSE MCP servers to every user at once, using the same entry format as .mcp.json. Entries that specify a command to run are skipped for safety.

What changes Instead of configuring MCP servers for each team member individually, an admin can now set them up once for everyone.Source →

Claude Code 2.1.259 Claude Code Adds --permission-prompts none for Unattended Runs

Added --permission-prompts none for unattended headless hosts: anything that would normally prompt for confirmation is now automatically denied, while the active permission mode, including auto mode, keeps deciding everything else.

What changes When running Claude Code in a CI pipeline or on a server where no one can answer a prompt, this keeps the run from hanging on a confirmation it can never receive.Source →

Claude Code 2.1.259 Claude Code Now Recognizes GitLab MR Commands

Claude Code now recognizes glab mr create/merge/close/reopen/note/update commands, showing them as "MR !N" in the collapsed tool summary and refreshing the footer MR badge.

What changes If you work in GitLab instead of GitHub, merge request commands now show up in the interface just as clearly as GitHub PR commands already did.Source →

Claude Code 2.1.259 Claude Code's Plugin Validator Gets a --json Output Option

Added --json to claude plugin validate, producing a machine-readable validation report.

What changes If you want to wire plugin validation into a CI pipeline to automatically pass or fail a build, you no longer need to parse human-readable output to do it.Source →

Claude Code 2.1.259 Claude Code Speeds Up Terminal Resize and First Render for Long Responses

Claude Code now reuses text measurements, improving terminal resize and first-render performance for long responses.

What changes If you keep several long responses open and resize your window often, the stutter that used to cause is now reduced.Source →

Claude Code 2.1.257 Claude Code's default model switches to Fable 5.1

Claude Code's default model is now Fable 5.1 (claude-fable-5-1), with a 1M-token context window, priced at $10/$50 per million input/output tokens and $0.25 per million tokens for cache reads.

What changes You'll be working on the cheaper, faster model starting your next session with no setup needed, and the savings are biggest on long sessions that lean on cache reads.Source →

Claude Code 2.1.257 Claude Code adds a time format setting

A new timeFormat/timeZone setting lets you choose 12-hour, 24-hour, 24-hour UTC, or a custom format for turn-end and transcript timestamps.

What changes Useful if you work across time zones or share logs with a team and want timestamps in a familiar format.Source →

Claude Code 2.1.257 Claude Code adds a containment-escape guard to auto mode

In auto mode, attempts to fetch cloud metadata credentials, evade egress monitoring, or reach across tenants are no longer auto-approved, unless your environment explicitly marks them as expected.

What changes If you let auto mode run unattended, risky actions like leaking cloud credentials are less likely to slip through without a human checking first.Source →

Claude Code 2.1.257 New env var forces one model across all subagents

Setting CLAUDE_CODE_SUBAGENT_MODEL_FORCE applies CLAUDE_CODE_SUBAGENT_MODEL (or the main model) to every subagent, overriding any per-spawn or agent-definition model choice.

What changes If running many subagents with mixed models made cost or behavior hard to predict, this one variable pins them all to a single model.Source →

Claude Code 2.1.257 /effort can now change effort for just the current session

Adding s to /effort now changes the effort level for only the current session, matching how /model already works, and it resets on the next session.

What changes Handy for trying a different effort level in just one session without changing your default going forward.Source →

ChatGPT ChatGPT for Clinicians adds US public healthcare data search

Eligible US users of ChatGPT for Clinicians can now search nine public healthcare data sources, including clinical trials, medication information, Medicare data, and provider records, directly inside ChatGPT.

What changes When you need to look up a clinical trial or drug information mid-consultation, you can check it inside ChatGPT instead of switching to another site.Source →

Claude app Claude app now runs on Fable 5.1 and Mythos 5.1

The Claude app's default models are now Fable 5.1 and Mythos 5.1.

What changes You don't need to change any settings, your next conversation already runs on the cheaper, faster model.Source →

Claude Code 2.1.251 Claude Code Cuts CPU Usage During Turns

Claude Code improved CPU usage during turns in interactive sessions by cutting redundant UI re-renders.

What changes Running several long sessions at once now spins up your laptop fan less and drains less battery.Source →

Claude Code 2.1.251 Claude Code Cloud Sessions Now Name the Cause of Connection Drops

When a cloud session's network proxy drops a connection during a Bash command, the tool result now names the host and reason instead of only showing 'connection reset.'

What changes When you're running several long sessions at once, you'll spend less time hunting for which window dropped and why.Source →

Claude Code 2.1.251 /schedule Now Explains Why It Can't Attach Your MCP Servers

MCP servers configured in Claude Code can't be attached to cloud routines. Instead of a bare 'No MCP connectors' message, /schedule now explains why.

What changes You'll spend less time confused about why a scheduled routine won't pick up your MCP connector, since the reason is now spelled out.Source →

Claude Code 2.1.251 Claude Code Now Tells Claude Its Subagent's Messages Aren't From an Outside Session

Claude Code now tells Claude that a message from its own subagent comes from a worker inside this session, not from an unrelated Claude session.

What changes When you run multiple subagents at once, Claude now responds with a more accurate sense of where each message actually came from.Source →