Claude Code defaults to Auto Mode on Pro, Max, Team starting Aug 14
A classifier that caught 89% of risky commands now gates them for you instead of endless approve prompts, and its overhead is now free on paid plans.
Copy markdown
Auto Mode becomes the default
From Aug 14, new Claude Code sessions on Pro, Max, and Team launch in Auto Mode: Claude runs commands on its own and a classifier only stops to ask when an action looks irreversible, destructive, or aimed outside your environment. Per-command approval prompts are no longer the starting point.
The 89% vs 13.6% pitch
Anthropic's case: across testing with 1,053 paid users, the classifier caught 89% of dangerous commands while human reviewers caught just 13.6% — because people approve about 97% of prompts reflexively. Teams on Auto Mode also opened roughly 25% more pull requests.
You can still take the wheel
Nothing is locked in. Switch modes anytime with Shift+Tab or the dropdown, and any default you've already set is kept. Auto Mode falls back to manual approvals after 3 blocks in a row or 20 in a session, and admins can pin an org-wide default or disable it entirely via managed settings.
The classifier overhead is now free
The tokens the safety classifier burns are no longer billed on Pro, Max, and Team — effective immediately, not Aug 14. So the more-autonomous default won't quietly eat into your usage limits.
Enterprise and API are next
Enterprise, API, and cloud-partner deployments stay opt-in for now, with Anthropic planning to flip their defaults within a month. If you run Claude Code through Bedrock or Vertex, expect the same change to reach you later.
Injection resistance, by the numbers
Third-party firm Trajectory Labs reported zero successful prompt injections in 720 attempts against Claude in Auto Mode, versus a 5.83% success rate against GPT-5.6 Sol. Anthropic still recommends reviewing Claude's actions yourself before it touches production infrastructure.