AI product notes: week of 17 August 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 v2.1.236: ANTHROPIC_DEFAULT_MODEL env var, cross-session idle notifications, and macOS sandbox wildcard-deny fixes (23 August 2026). The macOS sandbox fix closes a bypass where renaming a denied file could let sub-agents read protected paths, teams using wildcard read-deny rules in shared-repository workflows should update immediately.

• Anthropic: Claude Code v2.1.239, v2.1.241: data-residency cost transparency, /claude-api upgrade helper, and macOS sandbox tightening (23 August 2026). Teams running Claude Code in data-residency workspaces will now see the 1.1× inference surcharge surfaced in every cost estimate and budget flag, making it harder to accidentally overspend, and the /claude-api upgrade command makes the anthropic SDK 0→1.x migration a single command.

• Anthropic: Claude Enterprise Admin API user-management endpoints are now generally available (22 August 2026). Enterprise IT teams can now remove the beta header from their Claude user-provisioning scripts and treat member, invite, group, and custom-role management as stable, supported API surface.

• Microsoft: Microsoft Copilot Surveys agent retiring 31 August 2026: survey features move to Microsoft Forms (22 August 2026). Any Copilot workflow that uses the Surveys agent will stop working on 31 August, teams have nine days to rebuild those survey flows inside Microsoft Forms before the Agent Store listing disappears.

• OpenAI: ChatGPT improves plugin discovery with usage-ranked recommendations on web and mobile (22 August 2026). Plugin suggestions in ChatGPT now surface tools that users genuinely retain rather than ones with high install-but-abandon rates, which should make first-contact plugin discovery more useful across web and mobile.

• Anthropic: Claude Code adds ANTHROPIC_DEFAULT_MODEL env var, cross-session idle notifications, and tighter macOS sandbox (21 August 2026). Teams running Claude Code in CI or multi-session setups can now pin a default model org-wide via an environment variable and get notified when a long-running session goes idle, reducing the need to poll or watch logs manually.

• Anthropic: Claude Code v2.1.237: Concise output style and prompt-caching fix for LLM gateways (20 August 2026). Teams routing Claude Code through an LLM gateway should upgrade immediately to restore prompt-caching cost savings; the new ANTHROPIC_DEFAULT_MODEL variable is useful for CI pipelines that need a predictable starting model without overriding per-session choices.

• OpenAI: ChatGPT adds restaurant reservation search via OpenTable, Resy, and Yelp (19 August 2026). ChatGPT’s action surface now extends into live restaurant booking, adding another real-world task that can be completed end-to-end from a conversation rather than switching to a separate app.

• Anthropic: Claude Code v2.1.236: ANTHROPIC_DEFAULT_MODEL env variable, cross-session idle notifications, and macOS sandbox hardening (19 August 2026). CI and multi-agent pipelines can now pin a default model at the environment level without touching session configs, and macOS users get tighter sandbox defaults that reduce the blast radius of a runaway agent.

• Google: Gemini Enterprise Agent Platform Memory Bank IngestEvents API reaches general availability (19 August 2026). Agent builders who previously had to batch or manually trigger memory generation can now stream events continuously and control memory consolidation timing independently, enabling more responsive long-running agents.

• Anthropic: Gemini Robotics ER-1.6 shutdown aside: Anthropic biology safeguards updated in August, cutting Fable 5 biology fallbacks by ~85% (19 August 2026). Life-sciences developers and researchers using Claude Fable 5 through the API should see far fewer unhelpful refusals on legitimate biology questions, while the most sensitive autonomous-research scenarios remain gated behind Mythos 5 access controls.

• Google: Gemini-powered Admin Assist launches in Google Admin Console with Sidepanel and Search Overviews (19 August 2026). Workspace admins managing complex domain policies can now query documentation and troubleshoot settings conversationally inside the Admin Console, reducing the tab-switching overhead that currently slows incident response.

• Anthropic: Anthropic outage on 16 August 2026: authentication failure hit Claude.ai, Claude Code, and Claude Cowork for roughly 42 minutes (18 August 2026). The outage reinforces the ongoing reliability pattern that enterprise buyers need to account for when designing Claude-dependent workflows, especially agentic pipelines where mid-session authentication failures can corrupt in-progress tasks.

• OpenAI: OpenAI confirms ChatGPT Linux preview in its weekly roundup alongside the Copilot end-of-life confirmations (18 August 2026). The Copilot consumer feature removals are now complete with no data recovery path; the weekly note confirms there is no grace extension.

• OpenAI: OpenAI secures 8 GW at PORTS-Pike campus in Ohio and commits $84 M in Codex credits for students (17 August 2026). This is one of the largest single AI infrastructure commitments announced in the US, and the attached student Codex credit programme means hundreds of thousands of Ohio college students will get hands-on API access starting this academic year.