Week 39, 2026
Coverage window 18–24 September 2026.
For the interpretation of this week, and why an approval that expires matters more than a model that costs less, see the companion essay: Weekly AI Tools That Matter to Me: Part 9, Week 39. This page is the record; the essay is the read.
This week’s record
Three products from two vendors put an expiry date or a new owner on standing access, while two defaults at a third vendor moved towards fewer prompts.
- Claude Desktop v2.7032.0 added
mcpScheduledTaskApprovalLifetimeDays, which limits how many days a scheduled task may reuse a lasting MCP-tool approval before asking again. - Google Workspace added temporary administrator roles that are revoked automatically when they expire, and Gemini Enterprise added admin transfer of shared agents, which disables the agent’s schedules and triggers until the new owner turns them back on.
- ChatGPT changed new Health connections to default to “Allow low-risk actions” (14 September, a late capture), and announced that ChatGPT for Word will be enabled by default for Enterprise and Edu from 1 October.
- Anthropic released Claude Opus 5.5 and OpenAI released GPT-6 Sol and Luna on 22 September, both priced well below their predecessors.
- Antigravity 2.16 and 2.17 made plan review a named policy that can be set to skip review entirely, and stopped prompt actions inside widgets and HTML artifacts from sending automatically.
- Microsoft made federated Copilot connectors generally available, connecting Microsoft 365 Copilot to third-party sources over MCP under the user’s identity.
What changed this week
| Platform | What changed | Source |
|---|---|---|
| Anthropic ClaudeGA (Sep 22)Model release | Claude Opus 5.5: performs at the level of Fable 5.1 on most work and costs 40% less to run than Opus 5; $4/$20 per million input/output tokens, cache reads $0.20; first release since Anthropic called for pacing the frontier; externally evaluated before release; no longer offered with thinking switched off; Sonnet 5.5 and Haiku 5.5 to follow | Anthropic |
| Anthropic Claude DesktopReleased (Sep 22)Admin control | v2.7032.0: mcpScheduledTaskApprovalLifetimeDays caps how many days a scheduled task in Cowork or Code may reuse a lasting MCP-tool approval (0 removes the lasting option); inferenceCredentialKind: "external-idp" sends the user’s own OIDC token as the Bearer credential on every inference request; keepAwakeEnabled; OAuth issuer-mismatch validation for custom connectors enforced again | Changelog |
| Anthropic Claude DesktopFixed (Sep 22)Security fix | v2.7032.0 fixes: Microsoft 365 connector sign-in through the Windows broker requested only offline_access instead of the configured scope; managed connectors with their own OAuth client connected without credentials when the server allowed it | Changelog |
| Anthropic Claude DesktopReleased (Sep 21)Product release | v2.2553.13: Opus 5.5 support; Claude Code reads AGENTS.md where no CLAUDE.md exists | Changelog |
| Anthropic Claude DesktopReleased (Sep 17, late capture)Admin control / security fix | v2.2553.0: allowedPluginMcpServers (an empty list allows no plugin MCP servers); builtinBrowserEnabled with default, allowed and blocked domain policies; fixes for a routine with every connector removed still using all connectors, and for Plan mode skipping permission prompts on a fallback CLI | Changelog |
| Google AntigravityReleased (Sep 22)Product release / breaking change | 2.17.0: /plan with a Plan Review Policy of review every plan, review when the agent judges it worthwhile, or skip review; rules get a 20,000-token budget; per-project config moves to /.gemini/config.json and .agents/settings.json is no longer read | Changelog |
| Google AntigravityReleased (Sep 18–22)Product release / security | 2.16.0: subagents as live cards with a stop control; prompt actions in widgets and HTML artifacts fill the message box instead of sending; fix for reopening an older conversation re-applying its earlier file edits over newer work. 2.15.1 (Sep 19): file and network sandboxing on Windows. 2.15.0 (Sep 18): custom agents can switch off default prompts and tools | Changelog |
| Google Gemini EnterpriseAvailable (Sep 21)Admin control | Transfer ownership of shared agents: administrator-only, shared agents only, single owner; previous owner kept as agentUser; schedule and event triggers are disabled on transfer until the new owner enables them | Release notes |
| Google Gemini EnterpriseMixed, as statedProduct release | Gemini 3.8 Flash GA and enabled by default (Sep 23); D&B Risk Analytics data store GA with actions (Sep 22); new actions in Preview for OneDrive, Outlook, SharePoint and Teams, including sharing files and updating chat messages (Sep 18) | Release notes |
| Google WorkspaceAvailable (Sep 23)Admin control | Temporary administrator roles: roles assigned to users, groups or service accounts for a set period, revoked automatically at expiry; not assignable to the primary admin | Workspace Updates |
| Microsoft 365 CopilotGA (Sep 23)MCP | Federated Copilot connectors to third-party sources over MCP, data staying in the source system and accessed with user identity and admin oversight | Release notes |
| OpenAIGA (Sep 22)Model release | GPT-6 Sol and GPT-6 Luna: API prices 50% below GPT-5.6 promotional pricing (Sol $2/$10, Luna $0.10/$0.50); in ChatGPT Work and Codex for paid plans, not yet in Chat. Better prompt caching for GPT-6 with a dashboard, diagnostics and explicit breakpoints | GPT-6 Sol and Luna · Prompt caching |
| OpenAIPublished (Sep 22)Safety | Priorities and principles for effective third party assessments: four priority areas and definitions of safety claim and safety case | OpenAI |
| OpenAI ChatGPTRolling out (Sep 23)Product release | Voice with plugins, and Voice in Work; for Enterprise, “Enabling Voice does not enable Work or a blocked plugin”. Privacy Center (Sep 21) | ChatGPT release notes · Enterprise and Edu notes |
| OpenAI ChatGPT Enterprise/EduReleased (Sep 17, late capture)Admin control | ChatGPT for Word, with admin on/off; “Starting Oct 1, 2026, Word access will be enabled by default.” Tenant-wide SCIM extended to API Platform organisations and projects | Enterprise and Edu notes |
| OpenAI ChatGPTChanged (Sep 14, late capture)Default change | New Health connections default to the global plugin permission, “Allow low-risk actions”, using connected Health data without asking each time except for sensitive actions | ChatGPT release notes |
| OpenAI ChatGPTAnnounced (Sep 11, late capture)Deprecation | Planned retirement of custom GPTs with a migration path to plugins | ChatGPT release notes |
Late captures in this edition. Claude Desktop v2.2553.0 (17 September) and the OpenAI help-centre items dated 11 to 17 September fall inside Week 38’s window. The OpenAI release-notes pages had not been read in full since August; they are now read every edition.
No changes this week (scanned, nothing net-new)
- Anthropic corporate: beyond Opus 5.5, the window carries a science post (23 September) and an Accenture partnership (18 September).
- chiefmartec: no dated post found.
- Marketing AI Institute: practice articles only.
- Microsoft Agent 365: connected agents via A2A still not announced as generally available.
- Gemini Enterprise connector write actions: still in Preview.
Watching next (Week 40)
- ChatGPT for Word enabled by default for Enterprise and Edu on 1 October.
- Claude Sonnet 5.5 and Haiku 5.5, stated as following in the coming weeks.
- Custom GPT retirement dates and the migration path to plugins.
- Claude Desktop client-certificate fallback: in-app warning from 3 November, acceptance ends 17 November.
- Whether other vendors add expiry to standing agent approvals.
How this is compiled. Each edition is assembled from official release notes, changelogs and newsrooms. An AI assistant does the first-pass scan, reading release notes in a browser or through vendor-published Markdown routes, and I review and edit before publishing. Items are checked against the primary source; anything I could not confirm is marked as such. Dates and statuses reflect what was published as of the current-as-of date and can change.
Corrections and method notes. Spot an error? Tell me and I will fix it in the next edition. This edition logs late captures from Week 38’s window, most of them from OpenAI’s help-centre release notes, which the previous edition did not read.
AI Watch is compiled and edited by Andrea Veggiani, Head of Cross-Channel Marketing & MarTech Solutions at BitBang.