CONFIRMED: Claude Code 2.1.278 ends the hidden charge for the auto mode classifier

Anthropic released Claude Code 2.1.278 in the early hours of September 19, Brasília time, with a narrow but meaningful change for anyone already running auto mode. Starting with this version, the safety classifier that approves or blocks risky actions such as shell commands and network requests runs server-side for API, Enterprise, and Bedrock, Vertex, Foundry, and gateway deployments, at no charge for that overhead. Previously, every safety check consumed tokens like a normal model call, inflating the bill for anyone using auto mode heavily.

How the classifier fits into auto mode

Auto mode exists to cut down on the manual approvals a user has to grant during a Claude Code session. To do that, a classifier evaluates every potentially risky action, like running a terminal command or making a network request, and decides whether the task can proceed without human intervention. This process is what makes headless sessions and long-running agent workflows practical without constant supervision, but until now every classifier call was billed as part of the session's model usage.

What actually changes

With the classifier now running server-side as part of the session's own model requests, Anthropic stops charging for that overhead. The company kept an opt-out for anyone who prefers the old behavior: the CLAUDE_CODE_AUTO_MODE_SERVER=0 environment variable disables this routing on Bedrock, Vertex, Foundry, and gateways. When the server-side classifier cannot reach a user's session, the system falls back to the old, billed path and shows a clear warning so nobody is surprised by the invoice. The release also adds an 'Auto mode server' row to the /status command, explicitly showing whether that session is using the free classifier or has fallen back to the billed one.

The fine print that matters

Anthropic's official documentation spells out a practical condition: the benefit only applies from Claude Code 2.1.278 onward and depends on the session being eligible for server-side checks. Direct connections to the Anthropic API do not read the opt-out variable, and sessions that do not meet the eligibility criteria still see the standard billing notice. In short, this is not a blanket, automatic cost cut; it is a billing fix that depends on configuration and on the type of connection a company uses to run Claude Code.

What it means for a Brazilian SMB

For a small or midsize business already running Claude Code in auto mode inside CI pipelines, support agents, or internal automation, this change puts money back in the budget without requiring any rewrite. The gain is largest for anyone using auto mode intensively, with many shell and network actions per session, since that is exactly where classifier overhead weighed most on the bill. It is worth checking /status after updating to confirm the session is actually using the free classifier, especially in environments with a custom gateway or Bedrock, Vertex, or Foundry infrastructure, where eligibility is not automatic.

Trust maintenance, not hype

There is no new model or flashy feature in this release, and that is exactly why it deserves attention. Charging for a safety check the product itself requires to run without friction was always a quiet point of friction in auto mode. Fixing it increases cost predictability for anyone treating Claude Code as part of a production pipeline rather than a terminal toy. It is the kind of adjustment that never makes a headline but decides whether a team trusts an agent to run unsupervised all day.

Sources

Anthropic's official GitHub release: https://github.com/anthropics/claude-code/releases/tag/v2.1.278 | Official documentation on classifier billing: https://code.claude.com/docs/en/auto-mode-classifier-billing | Structured release record: https://api.github.com/repos/anthropics/claude-code/releases/tags/v2.1.278