AI product notes: week of 31 August 2026
A weekly roundup of smaller AI product updates, release-note items, and admin changes worth tracking.
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 v2.1.262: bug fixes and reliability improvements (6 September 2026). This is a maintenance-only Claude Code release; no action is required unless you were seeing the specific terminal or sandbox bugs it fixes.
- Google: Gemini in Workspace gains persistent custom instructions for all Ask Gemini surfaces, rolling to Scheduled Release tenants from 5 September (5 September 2026). Enterprise Workspace users on Scheduled Release can now save standing instructions, preferred tone, formatting rules, domain context, once and have Gemini apply them across every Workspace app, removing the need to re-prompt on every session.
- Google: Gemini Notebook audit logs now available in the Workspace Admin console and security investigation tool (5 September 2026). Admins in regulated industries now have a native audit trail for Gemini Notebook usage, covering who accessed what and when, which was a compliance gap since NotebookLM rebranded to Gemini Notebook in July.
- Google: Google Meet hardware devices can now join Microsoft Teams meetings directly (5 September 2026). Organisations running Google Meet room hardware no longer need a separate Microsoft Teams device to join Teams calls, a meaningful cost and complexity reduction for enterprises that use both platforms.
- Anthropic: Claude Code fixes macOS 12 Monterey launch failure and remote session permission bug (5 September 2026). Teams running Claude Code on macOS 12 Monterey were hit by a silent launch regression; this patch restores functionality without requiring an OS upgrade.
- OpenAI: Codex fixes GPT-6 Astra visibility in the bundled model picker and sets it as the bundled default (5 September 2026). Developers who rely on Codex’s bundled default will now automatically receive GPT-6 Astra for new sessions without any configuration change, but teams that want to pin to an earlier model should set an explicit model ID to avoid unintended upgrades.
- Anthropic: Claude Code ships fullscreen diff panel, prompt-cache miss diagnostics, and a batch of gateway and session fixes (4 September 2026). The new /diff panel gives a live view of file changes as Claude edits, and the cache-miss diagnostics make it much easier to understand why prompt caching is not saving tokens, useful for teams actively tuning agentic workflow costs.
- Anthropic: Anthropic launches Claude Academy to scale AI education for enterprise employees (3 September 2026). Enterprise learning and development teams evaluating AI upskilling programmes now have an Anthropic-built curriculum reference with structured frameworks for onboarding employees to safe AI use.
- Anthropic: Anthropic publishes prompt-injection evaluation results for Claude Cowork, showing near-zero attack success rates on current models (3 September 2026). Enterprise security teams evaluating Claude Cowork for agentic deployments now have Anthropic’s own quantified red-team numbers, showing Opus 5 in the Cowork harness resists professional prompt-injection attacks at a 3.8% pre-safeguard rate, the lowest published figure for any Claude model in this context.
- Anthropic: Claude Tag data-analytics public beta: a governed analytics agent that answers data questions directly in Slack (2 September 2026). Business teams can now self-serve data questions in Slack without a data-engineering handoff, but admins should review the permission and telemetry settings before enabling it organisation-wide.
- OpenAI: ChatGPT Work suffered elevated errors and latency on August 31 affecting Plus subscribers (31 August 2026). ChatGPT Work users, particularly Plus subscribers, faced hours of degraded or blocked task execution on August 31; enterprise teams running automated Work tasks should review incident logs and check whether any scheduled tasks need to be rerun.
- Microsoft: Microsoft’s August 2026 Copilot roundup: Cowork effort controls, GitHub Copilot credit tracking in Insights, and Planner agent integration coming in September (31 August 2026). Cowork’s new effort slider lets users trade speed for depth on a per-task basis, and admins can now see GitHub Copilot credit burn alongside other AI spend in a single Insights dashboard, both changes affecting how teams budget and govern their AI usage.
- OpenAI: OpenAI supports California’s youth AI safety bill (31 August 2026). OpenAI’s endorsement of California’s youth AI safety bill signals the company is moving toward embracing, rather than resisting, state-level AI regulation, which could set a precedent that shapes compliance requirements for AI products used by minors across the US.