AI product notes: week of 14 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

• Anthropic: Claude Code v2.1.275, v2.1.278: gateway sign-in account confirmation, skills and plugins sync from claude.ai, and send-now key (20 September 2026). The gateway account confirmation closes a sign-in spoofing gap for enterprise proxy setups, while the automatic claude.ai skills/plugin sync means terminal users now share the same toolset as their web sessions, unless they explicitly opt out.

• Google: Gemini API releases Antigravity Agent 09-2026, replacing antigravity-preview-09 (20 September 2026). Developers building on the Antigravity preview endpoint need to migrate to the new antigravity-agent-09-2026 model ID before the preview is retired, and can now ship agents that suppress Antigravity’s default prompts and tools for a cleaner user experience.

• OpenAI: Codex TUI local sessions fix: reasoning summaries now off by default for providers that don’t support them (20 September 2026). Developers who use Codex locally via the TUI with non-OpenAI providers, or API-compatible proxies that don’t support reasoning summaries, will no longer see request failures after updating.

• OpenAI: ChatGPT Computer History expands to EEA, Switzerland, and UK for Pro users (19 September 2026). Pro users in the EEA, Switzerland, and UK can now enable Computer History to let ChatGPT resume tasks across their Mac sessions with full local-timeline context, though the feature remains opt-in and stores data only on-device.

• OpenAI: ChatGPT desktop app gains quick-chat floating control and Windows appshots in latest release (19 September 2026). Power users and developers who run ChatGPT alongside other apps can now fire off a quick prompt from a floating control without interrupting their current workflow, and Codex Micro lets agents insert generated text directly into any app on the desktop.

• OpenAI: ChatGPT Health plugin connections now default to your global plugins setting, and Thinking auto-switching retires for Plus and Pro (18 September 2026). Plus and Pro users who relied on automatic Thinking escalation for complex health queries will now need to select the model manually, and admins managing Health plugin rollouts should note the default now follows global plugin permissions rather than a per-plugin toggle.

• Google: Workspace Studio Flows admin controls for custom steps, integrations, and webhooks begin rolling out 17 September 2026 (18 September 2026). Admins who want to govern which external endpoints their users’ Workspace Studio flows can call need to configure allowlists before end-user features go live on 21 September 2026, otherwise flows will be able to post to any valid URL.

• Anthropic: Claude Code fixes read-only git commands asking for permission after extended sessions (regression in 2.1.269) (18 September 2026). Long-running Claude Code sessions on 2.1.269 were being interrupted by spurious permission prompts on read-only git operations; updating to the latest release removes the interruption without changing your permission settings.

• Anthropic: Claude Code changelog: latest release on 17 September 2026 (17 September 2026). Developers running Claude Code in CI or on Max/Enterprise plans should check the 17 September changelog for any breaking changes or new defaults in the latest build before updating.