Week 40, 2026
Coverage window 25 September – 1 October 2026.
For the interpretation of this week, and why an agent with its own identity changes what an audit log can tell you, see the companion essay: Weekly AI Tools That Matter to Me: Part 10, Week 40. This page is the record; the essay is the read.
This week’s record
OpenAI introduced dots, always-on agents that work on the user’s behalf, and previewed specialist dots with their own identity for access management, while Anthropic, Google and OpenAI’s own enterprise controls tightened what an agent may do under a person’s identity.
- OpenAI held DevDay on 29 September: dots, GPT-6.1 Sol, an Ultrafast tier, Codex Cloud, plugin extensions, MCP Events, ChatGPT Space, Pages, Teams, @ChatGPT in Slack and Microsoft Teams, and an OpenAI Marketplace in beta.
- Claude Desktop made an organisation’s “Restrict to Ask” connector policy hold even in Skip all approvals mode, and added a deny list for plugin MCP servers.
- Antigravity fixed a defect that let
.git,.envand.vscodefiles be modified without a permission prompt, and added a plugin marketplace. - Anthropic released Claude Sonnet 5.5, the first Sonnet launched with cyber safeguards that fall back to Sonnet 5, and Google announced Gemini 4 Argon, rolling out first to trusted cyber defenders.
- Google set dates for replacing Gems with skills built on the SKILL.md standard, and let the Docs API submit suggested edits instead of changing content directly.
What changed this week
| Platform | What changed | Source |
|---|---|---|
| Anthropic ClaudeGA (Sep 28)Model release | Claude Sonnet 5.5: “runs 30%+ faster, and costs up to 30% less for most work” than Sonnet 5, at the same $2/$10 per million input/output tokens; first Sonnet launched with cyber safeguards, where higher-risk cybersecurity tasks “will visibly fall back to Sonnet 5”; classifiers preventing reasoning extraction; thinking-off users move to the between_tools setting; Haiku 5.5 “in the coming weeks” | Anthropic |
| Anthropic Claude DesktopReleased (Sep 29)Admin control / security fix | v2.16120.0: an organisation’s “Restrict to Ask” connector policy now asks “including in ‘Skip all approvals’ mode” (newly started tasks); deniedPluginMcpServers blocks plugin MCP servers by URL even when the allow list admits them; userPluginMarketplacesEnabled: false refuses installs and updates from unprovisioned marketplaces; fix for proposed skills saved empty | Changelog |
| Anthropic Claude DesktopReleased (Sep 27)Product release | v2.9939.4: Claude Sonnet 5.5 support | Changelog |
| Anthropic Claude DesktopReleased (Sep 24, late capture)Default change / admin control | v2.9939.2: unless autoModeEnabled is false, new Code tab sessions start in Auto mode where the model supports it; managed MCP server connections no longer follow HTTP redirects; connectors start disconnected when a different person signs in to a shared computer login; deleting a Code session removes its transcript | Changelog |
| GoogleLimited rollout (Sep 30)Model release | Gemini 4 Argon: frontier model for long-horizon engineering, legal, finance and cyber-defence work; rolling out first to trusted cyber defenders through the Fairwind Program; introductory price $2/$10 per million input/output tokens | |
| Google AntigravityReleased (Sep 28)Product release / security fix | 2.18.1: Customizations tab with a plugin marketplace; syntax-highlighted commands in permission requests; fix: files in .git, .env and .vscode directories could be modified without a permission prompt under the Default and Request Review presets | Changelog |
| Google AntigravityReleased (Sep 30)Product release / fix | 2.19.1: send messages directly to a subagent; export rendered Markdown as PDF; conversation-only undo; fix: custom agents ignored global and project rules | Changelog |
| Google Gemini EnterpriseAvailable / PreviewAdmin control | Granular resource-level IAM for individual apps and data stores, with custom roles for restricted administrators, viewers and end users (Sep 28). New Outlook actions in Public Preview, including forward, move and reply; new Preview data stores including Zapier (Sep 29) | Release notes |
| Google WorkspaceAnnounced (Sep 30)Deprecation / product | Skills replace Gems: built on the “open, Markdown-native SKILL.md standard”; Workspace rollout from Oct 5, Gemini app from Oct 13; Gems stop working for business and enterprise no sooner than 1 March 2027, with auto-migration to draft skills | Workspace Updates |
| Google WorkspaceRolling out (Sep 30)API / agent pattern | Docs, Sheets and Slides APIs support programmatic comments, and the Docs API supports suggested edits, letting automated tools “propose document revisions for user review rather than altering content directly”; MCP-server comment support remains in developer preview | Workspace Updates |
| OpenAIRolling out (Sep 29)Agentic AI | dots: always-on agents on GPT-6 Astra with their own cloud computer and browser, working on the user’s behalf, with app access and permissions managed through existing ChatGPT app controls; read-only background research, built-in and custom rules to allow, require approval for or block actions, and automatic verification of account-affecting actions. Specialist dots (preview, enterprise pilots) with their own identity, credentials and system access, with Microsoft Agent 365 integration. Pro and Business Premium rollout; Pro excludes the EEA, Switzerland and the UK at launch; Enterprise beta off by default | Introducing dots · ChatGPT release notes |
| OpenAIGA (Sep 29)Model release | GPT-6.1 Sol: near-Astra performance at a fifth of Astra’s standard prices; cached input $0.10 per million; Enterprise and Edu administrators must enable access. Ultrafast tier and Pro 500 plan | GPT-6.1 Sol · DevDay recap |
| OpenAIMixed, as stated (Sep 29)Product release | DevDay platform items: Codex Cloud, Codex Security Cloud, Decisions API (limited preview), computer use in the Agents API, Amazon Bedrock Managed Agents; plugin extensions and MCP Events for automations; ChatGPT Space, Pages, collaborative Slides, Teams and team tasks, @ChatGPT in Slack and Microsoft Teams; Sign in with ChatGPT; OpenAI Marketplace (beta) with partners including Adobe, HubSpot and Salesforce | DevDay recap |
| OpenAI ChatGPT Enterprise/EduReleased (Sep 29)Admin control | Team connections: “Sharing a plugin and granting access to an app connection are separate settings.” Sites with connected apps off by default per plugin; “Visitor access … can’t be reused by background scheduled tasks.” | Enterprise and Edu notes |
| OpenAI ChatGPT Enterprise/EduAdmin preview (Sep 24, late capture)Admin control | External access controls: whether Sites can use members’ connected apps and whether applications can access ChatGPT Ads; both off by default during the admin preview | Enterprise and Edu notes |
| OpenAIPublishedSecurity / safety | Security history for OpenAI accounts (Sep 25); disrupted a coordinated model distillation campaign (Sep 30); towards safety cases for frontier AI training (Sep 28) | ChatGPT release notes · OpenAI news |
Late captures in this edition. Claude Desktop v2.9939.2 (24 September) and ChatGPT Enterprise’s External access controls (24 September) fall inside Week 39’s window and were not in that edition.
No changes this week (scanned, nothing net-new)
- Anthropic corporate: beyond Sonnet 5.5, a Barclays customer story (1 October).
- chiefmartec: no dated post found.
- Marketing AI Institute: practice and event articles only.
- Microsoft 365 Copilot: no release-notes batch after 23 September. Agent 365 connected agents via A2A still not announced as generally available.
- Gemini Enterprise connector write actions: still in Preview.
Watching next (Week 41)
- dots availability in the EEA, and the usage terms OpenAI says it will publish after the first month.
- Claude Haiku 5.5, and the managed-configuration cut-off for older key spellings on 7 October.
- Skills rollout in Workspace from 5 October and in the Gemini app from 13 October.
- Gemini 4 Argon availability beyond the Fairwind Program.
- Custom GPT creation ending 26 October for affected Enterprise workspaces.
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. OpenAI’s newsroom pages redirected to the Italian locale for this edition; descriptions of dots are taken from the English launch page and help-centre release notes. The Italian page describes every dot as having its own identity for access and permissions; the English page reserves its own identity for specialist dots, and this edition follows the English page.
AI Watch is compiled and edited by Andrea Veggiani, Head of Cross-Channel Marketing & MarTech Solutions at BitBang.