Claude Code stops billing API users for the auto mode check

The interesting bit in v2.1.278 is a line item most people never saw: auto mode's routing decision is itself billable work. Per the release notes on Github, auto mode for Claude API and Enterprise users, and on Bedrock, Vertex, Foundry and gateways, now defaults to the server-side classifier, which does not charge for classifier overhead.

At a glance

• A new Auto mode server row in /status now tells you, for the current session, whether the auto mode classifier is running on the server or on the path you get billed for.
• The new default points at the server-side classifier for Claude API and Enterprise users and on Bedrock, Vertex, Foundry and gateways, where CLAUDE_CODE_AUTO_MODE_SERVER=0 opts out.
• Sessions can still fall back to the billed classifier, and Claude Code warns when that happens, but the notes do not say what pushes a session onto that path.

If you have not been following auto mode: it is the setting where Claude Code picks the model for your turn instead of you choosing by hand. Something has to make that pick, and that something is a classifier, a small fast judgement about what you just asked for. Running it costs tokens, and the release notes call those tokens classifier overhead.

What v2.1.278 changes is where that classifier runs by default. For Claude API and Enterprise users, and on Bedrock, Vertex, Foundry and gateways, auto mode now goes to the server-side classifier, with no charge for the overhead. On Bedrock, Vertex, Foundry and gateways you can opt out with CLAUDE_CODE_AUTO_MODE_SERVER=0. The notes point to a documentation page on auto mode classifier billing.

The other half of the change is visibility. /status gained an Auto mode server row that shows whether this session's auto mode classifier runs on the server. When a session does drop to the billed classifier, Claude Code warns you rather than leaving you to spot it later.

Mechanically the two paths differ only in who executes the judgement. On the server path the routing decision happens on Anthropic's side and never turns into tokens on your account; on the billed path the classifier call goes through the same metered pipe as your actual work. It is the difference between asking the receptionist which desk you need and paying a consultant to tell you.

What the notes do not spell out is the trigger for that billed fallback, or how much classifier overhead amounted to in the first place; no figure for it appears anywhere in the release. In our view the warning is the right call for a change like this, because you learn you are on the metered path during the session instead of afterwards in the usage report.

What the billing docs still owe

The open item is the fallback. Until the auto mode classifier billing page spells out what sends a session to the billed classifier, the /status row is the only dependable way to check where a given session sits. The notes also do not say whether the CLAUDE_CODE_AUTO_MODE_SERVER=0 opt-out applies anywhere beyond Bedrock, Vertex, Foundry and gateways, and they name no version where that would change.