Claude Code stops billing API users for the auto mode check

Claude Code 2.1.278, released on September 19, moves auto mode’s safety check onto Anthropic’s servers for Claude API and Enterprise users and on Bedrock, Vertex, Foundry and gateways. The release notes on GitHub say this server-side classifier is now the default for those users, and it doesn’t charge for classifier overhead.

Auto mode depends on the classifier. Every tool call passes through it, and it blocks anything irreversible, destructive or aimed outside the environment. Deny rules and explicit ask rules are checked before the classifier runs and still block or prompt. So the classifier’s cost comes with every action the agent takes, which is why it matters who pays for it.

Auto mode has been the default for new sessions on Pro, Max and Team plans since August 14. The Claude Code docs now list it as available on every provider, including the Anthropic API, Amazon Bedrock, Google Cloud’s Agent Platform and Microsoft Foundry. This release changes who pays for the check on those providers, which is where per-token billing applies.

Setting CLAUDE_CODE_AUTO_MODE_SERVER=0 opts out of the server classifier, but the release lists that switch only for Bedrock, Vertex, Foundry and gateways. It gives no equivalent for Claude API or Enterprise users. A session can also fall back to a billed classifier, and Claude Code shows a warning when that happens. It’s unclear what triggers the fallback. The billing rules are on a separate docs page.

You can check where a session’s classifier runs with /status, which now has an Auto mode server row. That row shows whether the classifier for that session runs on the server.