Code Review is now generally available for Claude Team and Enterprise plans. Reviews are faster than the research preview, catch more meaningful bugs, and admins can configure review depth (Quick, Standard, or Deep) per repository.
GA Code Review is 3x faster than the research preview, with higher-quality findings that include the problem, trigger, and fix in each comment. Admins can select Quick, Standard, or Deep review depth per repository.
Code Review has moved from research preview to GA for Claude Team and Enterprise plans. Reviews are now 3x faster, surface more meaningful bugs with concise findings, and admins can choose Quick, Standard, or Deep review depth per repo.
Admins enable Code Review in Claude Code settings and select repos. After that, reviews run automatically on every new pull request with no additional setup required from individual developers.
Sessions with 200K or more tokens that sit idle for about 54 minutes now auto-compact, so returning users get a warm summary instead of a cold re-cache when the prompt cache expires at the one-hour mark. The token threshold is configurable via CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS, and the feature can be disabled with 'idleCompaction': false.
Users constantly approving tool-use prompts can enable Auto mode to bypass the manual confirmation loop entirely.
From June to September, merged PRs more than doubled and the share of reverted PRs fell over 40%, driven by changes to how the team does code review with Claude. A blog post details the approach and how others can adopt it.
A member of the Claude Code team quotes the official stat: roughly 90% of Claude Code's own codebase is written by Claude, a detail that lands differently coming from someone on the inside.
Responding to a user who argued the model should just write correct code from the start, a team member acknowledged that's the eventual goal but called code review 'extremely useful in the meantime' while models keep improving.
Agents in a workflow inherit whatever model the session is using unless the script explicitly routes them elsewhere. The recommended pattern: plan with a capable model, then run /model sonnet before launching the workflow, or just tell the orchestrator to use Sonnet for build agents.
Responding to a user excited about automating their home heating backup, an Anthropic team member casually mentioned Opus 5.5 as an upcoming model that will make such projects easier.
An Anthropic team member is working on making their Claude-powered smart home repo public. It has been running for about a week and, per the author, outperforms both Apple Home and Home Assistant. The holdup is disentangling personal monorepo integrations.
A demo of a smart home app where Claude interprets plain-language descriptions of automations and configures the underlying rules automatically, no manual rule-building required.
Conversations sent to the model are now gzip-compressed on every turn, reducing latency. Cloud sessions also begin initializing the moment you start typing, so the first response arrives sooner.
In Mac Claude 2.37337.1, messages from orchestrator to child subsessions are queued until the end of a full run, meaning they arrive stale and only intermittently. A team member flagged the report as specific and helpful.
Responding to a question about who writes the remaining 10% of code not attributed to Claude, bcherny clarified it is also Claude. The team is tightening attribution confidence before claiming a full 100%.
