Anthropic will make “auto mode” the default way Claude Code runs for Pro, Max, and Team subscribers starting August 14, replacing the tool’s traditional manual-approval prompts, the company said in a blog post this week.
Auto mode routes every action Claude wants to take — running a command, editing a file, pushing to a repository — through a safety classifier instead of stopping to ask the user each time. The classifier is built to block actions judged irreversible, destructive, or reaching outside the current project, while allowing routine work to proceed uninterrupted. If it blocks something, Claude typically looks for another way to do the task or asks the user directly.
Why Anthropic made the switch
Anthropic says the case for auto mode rests on data, not convenience. In a study of 1,053 paid users, manual review by humans caught only 13.6% of commands later judged dangerous, while auto mode’s classifier caught 89% of the same set. Head-to-head, auto mode blocked roughly 800 risky commands that human reviewers had waved through, while humans blocked only six actions the classifier had allowed. Anthropic says the gap widens the longer a session runs, as human attention fades but the classifier’s detection rate does not.
A separate analysis of real production sessions from May and June found manual approval let unintended harm through more than twice as often as auto mode did, at 6.3% versus 2.4% of flagged sessions. Third-party red-teaming by Apollo Research and a 720-run prompt-injection test by Trajectory Labs are both cited as further validation; in the latter, auto mode blocked all attempted attacks against Claude’s Fable 5, Opus 5, and Sonnet 5 models.
What changes for users
For Pro, Max, and Team plans, new Claude Code sessions will default to auto mode from August 14, though anyone who has already set a manual-approval preference gets a one-time prompt rather than a silent switch. Anthropic is also dropping the extra token charge the safety classifier previously added for these plans. Enterprise customers, the Claude API, and cloud platforms including AWS, Google Cloud, and Microsoft Foundry keep auto mode as an opt-in for now, with Anthropic saying it plans to extend the default there within the coming month.
Anthropic reports that teams already using auto mode ship about 25% more pull requests than those relying on manual review, and cites customers including Adobe, Nuro, Gusto, and Garner Health describing longer, less-interrupted coding sessions as a result.
The move follows a string of recent changes to how Anthropic governs Claude Code’s autonomy, including a sharply trimmed system prompt for newer models and the addition of a dedicated security scanning plugin.