Claude Code stops charging API users for auto mode's safety checks, and now reads AGENTS.md

Claude Code has stopped billing part of its own safety machinery, but only for some accounts. Version 2.1.278, released on September 19, changes auto mode for Claude API and Enterprise users, and on Bedrock, Vertex, Foundry and gateways, to default to a server-side classifier "which does not charge for classifier overhead".

Auto mode runs a safety check before shell commands and network requests.

Auto mode is the permission setting that lets Claude Code run commands without stopping to ask each time. Anthropic's documentation says that in auto mode "a classifier runs safety checks on actions such as shell commands and network requests before they run", and until this build those checks were extra model requests that landed on the bill as token usage.

The server does the checking, when it can reach you

From 2.1.278, Anthropic says 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". The default applies to Enterprise plans, accounts on the Claude API, and sessions on Claude Platform on AWS, Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry, with the caveat that each platform and region gets it on its own rollout schedule.

When the server's checks cannot reach a session, nothing breaks and nothing changes: Claude Code goes back to running the classifier itself, and those requests are billed as before. The first time that happens, the session holds the action and prints a one-off notice: "We're changing auto mode to no longer charge for classifier requests in Claude Code. However, this session isn't eligible."

Pressing Enter runs the held action and the rest of the session on the billed classifier, and the notice does not come back in that session. Pressing Esc or Ctrl+C cancels the action and stops the turn, with auto mode still on. Anthropic says Pro, Max and Team plans never show the notice at all.

There is a way to check without waiting for a warning. The same release adds an Auto mode server row to /status, which reads Enabled while the server is deciding the session's actions and Disabled once it has fallen back.

A gateway is usually what breaks it

Anthropic names the likely culprit plainly. The most common cause is "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", which is a common thing to put in front of a model provider for logging and cost control.

The fix is on the gateway's side: pass the safeguards request field and the safeguard_results response field through untouched. Teams that know their gateway cannot do that can set CLAUDE_CODE_AUTO_MODE_SERVER=0 before starting a session, which keeps the billed classifier and silences the notice. Anthropic calls that variable "a temporary setting" that "may be removed in a later release", and says it is not read on a direct connection to the Anthropic API. A gateway is not the only cause. The docs say the notice can also appear when the platform, region or credential a session uses has not been given the server's checks yet.

Anthropic has published no figure for what the classifier overhead cost, so the change arrives without a number attached. It is a default rather than a price cut, since the old billing still applies whenever the server's checks go missing. It does land in the middle of a busy month for the product, which saw Claude Cowork folded into the Claude chat app on September 18.

Version 2.1.277 picks up AGENTS.md

The build a day earlier, 2.1.277 on September 18, added something smaller that will matter to anyone running more than one coding agent in the same repository. In a project with no CLAUDE.md, Claude Code now reads AGENTS.md instead, the cross-tool instructions file that Codex and several other agents use.

It is a fallback, not a replacement: an existing CLAUDE.md still wins, and the behaviour can be changed under "Project instructions" in /config. Anthropic's note says the feature is "not yet on Bedrock, Vertex or Foundry". For anyone who has been hunting for ways to bring a Claude Code bill down, the classifier change is the bigger of the two, and it arrives without anything to switch on.