AI product notes: week of 21 September 2026

Smaller product updates do not always need a full article, but they can still matter if you own the platform, run the admin console, or build against the API.

Updates

OpenAI: ChatGPT gets Apple Messages plugin, Computer History expands to EEA, and Sites URL renaming (27 September 2026). Apple silicon Mac users running Codex or ChatGPT Work can now let the AI read and send iMessages from within the app, a meaningful integration for personal and team communications workflows, while European Pro users finally gain the activity timeline feature that U.S. users have had since August.

Microsoft: Microsoft cancels two promised Copilot features: mobile push notifications and Interactive Agents for Teams Meetings (27 September 2026). IT teams and developers who planned workflows around Copilot mobile push alerts or interactive meeting agents should remove those features from their roadmaps, neither will ship.

OpenAI: ChatGPT adds Security History so users can audit sign-ins, MFA changes, and passkey events (26 September 2026). Users and IT teams can now review a timestamped log of every sign-in and security-setting change on a ChatGPT account, making it easier to detect unauthorised access and satisfy internal security audits.

OpenAI: Codex CLI 0.157.1 patch released on 26 September 2026 (26 September 2026). A patch release to Codex CLI landed on 26 September; teams running 0.157.0 on npm should update, but no documented functional changes are available from the official changelog.

Anthropic: Anthropic publishes Project Swap: Claude agents achieved 61% preference alignment negotiating book trades with minimal human input (25 September 2026). The experiment shows Claude agents can represent human preferences in real economic negotiations with minimal onboarding time, a signal that agentic commerce deployments may require less prompt engineering than previously assumed, while raising new questions about what agents do when instructions are ambiguous.

Anthropic: Claude API adds anthropic-workspace-id response header and MCP connector tool-list pinning via mcp-client-2026-09-15 beta (25 September 2026). API developers building multi-workspace or gateway setups can now reliably identify which workspace resolved a request via the new response header, while the MCP tool-list pinning beta reduces repeated server round-trips and makes tool availability deterministic across long agentic runs.

Anthropic: Claude Code September 25 release: gateway prompt-ID headers, model deny-lists, OTel tool logging, and /doctor prompt-audit (25 September 2026). Enterprise gateway operators gain per-prompt request grouping for cost attribution and debugging, while IT admins get granular model deny-listing to block specific Claude releases from ever running in their environment, two controls that were missing from the managed-settings surface.

Anthropic: Claude Code ships broad release with new terminal controls, MCP reliability fixes, and managed Chrome policy (25 September 2026). Enterprise Claude Code deployments gain a new managed policy to run Claude in Chrome alongside a locked-down MCP config, while all users benefit from faster startup, cleaner MCP error handling, and fixes to several background-agent reliability issues that caused sessions to stall or lose data.

Microsoft: Copilot+ PC branding quietly dropped from new Microsoft Surface and Qualcomm Snapdragon device line-up (25 September 2026). The retreat of the Copilot+ branding from new Surface hardware suggests Microsoft is distancing itself from the consumer AI device push that defined its 2024, 25 strategy, consistent with today’s broader pivot to enterprise-only Copilot positioning.

Microsoft: Word, Excel, and PowerPoint are now embedded directly inside the Copilot app with live data sync (25 September 2026). Users can now create, edit, and share Office documents without leaving Copilot, documents stay in sync with the rest of M365, and AI assistance is available inline, removing the context-switching that has been the main friction point in the current side-panel experience.

Anthropic: Claude Code adds Trusted Devices support for Pro and Max plans (25 September 2026). Pro and Max plan users running Claude Code across multiple sessions no longer need to re-authenticate on machines they have marked as trusted, reducing friction in long-running agentic workflows.

Microsoft: Microsoft 365 and Copilot X social accounts merge: all product updates moving to @MSFTCopilot (25 September 2026). Admins and developers who follow @Microsoft365 on X for Copilot release notes will no longer receive updates there; all Copilot product news is now consolidated at @MSFTCopilot, and the web home for the unified app is moving to copilot.cloud.microsoft.

Anthropic: Claude Compliance API now returns Claude in Chrome session transcripts for Enterprise beta users (24 September 2026). Enterprise compliance teams can now pull Claude in Chrome session transcripts alongside other Claude sessions through a single Compliance API call, closing the last major gap in audit coverage for organisations that allow browser-based Claude use.

Anthropic: Anthropic launches Claude for Small Business partner webinar series with 14 technology connectors, starting late September (24 September 2026). Small business owners evaluating Claude integrations can attend free partner webinars to see how specific connectors handle their data before committing, a lower-friction way to vet the tools that connect Claude to existing business software.

Anthropic: Anthropic signals Sonnet 5.5 and Haiku 5.5 are coming within weeks as the 5.5 family launches (22 September 2026). Teams running production workloads on Sonnet or Haiku can expect cheaper, faster 5.5-generation equivalents within weeks, worth pausing new Sonnet/Haiku architecture decisions until the new models land.

OpenAI: ChatGPT Health plugin connections now default to each user’s global plugins permission setting (22 September 2026). Health plugin users who connect a new service will now find it inheriting their global plugin permissions rather than starting locked down, which may automatically allow low-risk actions they haven’t explicitly reviewed.

OpenAI: ChatGPT now supports multiple accounts per plugin, starting beyond Gmail and Google Calendar (22 September 2026). Users who need ChatGPT to act across both personal and work accounts in the same conversation can now do so without switching sessions, reducing friction for people bridging consumer and professional tools.

Anthropic: Claude Code v2.1.280 upgrades Bedrock, Vertex, and Foundry installs to MCP 2026-07-28 by default (22 September 2026). Enterprise Claude Code installs running on Bedrock, Vertex, or Foundry will now use the current MCP protocol by default, teams with custom MCP servers that haven’t been updated to the July 2026 spec need to test or set the opt-out flag.

Microsoft: Microsoft 365 Copilot usage-based billing spending policies auto-apply to new services by default (22 September 2026). If your organisation has a Copilot spending policy, any new Microsoft agent or AI service that launches with usage-based billing will automatically start drawing from that budget unless an admin turns off the auto-apply setting per policy.

OpenAI: OpenAI expands Academy with role-based courses, assessments, and badges for developers, leaders, and educators (22 September 2026). Enterprise AI administrators and L&D teams now have a structured, vendor-provided curriculum they can point employees to, including verifiable badges, as organisations look to demonstrate responsible AI adoption.

Anthropic: Claude Code gains AGENTS.md support and MCP v2 for Bedrock, Vertex, and Foundry in latest release (21 September 2026). Teams using Claude Code on Bedrock, Vertex, or Foundry should verify their MCP server configurations are compatible with the v2 protocol, as the legacy path now requires an explicit opt-out.