Claude Code 2.1.278 stops charging for auto mode's safety checks, unless your gateway drops a field

Claude Code 2.1.278, published to npm on September 19, 2026, moves auto mode's safety checks onto Anthropic's servers and stops charging for them there. The changelog's first line:

"Changed auto mode for Claude API and Enterprise users, and on Bedrock, Vertex, Foundry and gateways, to default to the server-side classifier, which does not charge for classifier overhead (CLAUDE_CODE_AUTO_MODE_SERVER=0 opts out on Bedrock, Vertex, Foundry and gateways); warns on billed fallback."

Auto mode is the permission mode in which a classifier, rather than the user, approves actions such as shell commands and network requests. Until now those classifier calls were extra model requests on the user's bill. The new documentation page on classifier charges explains what changes and, more usefully, when it doesn't.

How it works

Per the docs, Claude Code "asks the server to perform those checks as part of the session's own model requests, and doesn't charge for them when the server performs them." When the server's checks can't reach a session, Claude Code falls back to its own classifier requests, "and those requests are billed as they were before."

The default applies to "Enterprise plans and accounts that use the Claude API," plus Claude Platform on AWS, Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry, with the caveat that "Whether a platform or region performs them depends on that platform's rollout." The page adds: "Pro, Max, and Team plans never show the notice."

The notice that holds your first action

The fallback is not silent. Before the first action it would check the billed way, Claude Code holds that action and shows:

We're changing auto mode to no longer charge for classifier requests in Claude Code. However, this session isn't eligible.

Enter continues on billed classifier requests for the rest of the session. Esc or Ctrl+C cancels the held action and stops the turn. In non-interactive runs with -p the text goes to stderr, and in stream-json output it arrives as a system warning message. A new Auto mode server row in /status reads Enabled or Disabled.

Headless pipelines should grep for that warning. It is how an unattended job reports that its classifier requests are billed again.

Gateways are the usual cause

The docs name the most common culprit: "an LLM gateway or proxy between Claude Code and the API: one that strips or rewrites request headers, drops request fields it doesn't recognize, or edits responses, for example by rewriting IDs or dropping keys from streaming events."

The fix belongs to whoever runs the gateway. It has to forward "request headers and body fields as they are, including ones the gateway doesn't recognize such as the safeguards request field," and return responses "without dropping keys such as the safeguard_results field or rewriting tool-use IDs."

If a gateway can't do that, CLAUDE_CODE_AUTO_MODE_SERVER=0 stops Claude Code asking, so the notice never appears and classifier requests stay billed. Two limits: the variable "isn't read on a direct connection to the Anthropic API," and it "is a temporary setting and may be removed in a later release." Setting CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 with the variable unset also turns the server's checks off.

What the docs do not say

The page says the server-side checks are not charged. It gives no figure for what the client-side classifier cost before, so any saving has to be measured against your own usage records. On Bedrock, Agent Platform, Foundry and signed-in gateway sessions, auto mode itself remains limited to "Claude Sonnet 5, Opus 4.7 or later, and the Fable models."

Primary sources: Claude Code CHANGELOG, Claude Code docs: Auto mode classifier request charges, npm registry: @anthropic-ai/claude-code, read 2026-09-22.