Claude Code 2.1.278 stops charging for auto-mode safety checks on Claude API and Enterprise plans
Why it mattersTeams that keep coding agents in auto mode were paying for extra classifier calls on every shell command and network request; that cost is gone for API and Enterprise plans on a direct path to Anthropic, and Claude Code shows in `/status` whether the current session is eligible.
Anthropic released Claude Code v2.1.278 on 19 September. In auto mode, the tool used to run a small classifier request against each shell command or network request it was about to make, and those requests counted as token usage on the bill. The new default sends those safety checks with the session's own model request, and Anthropic says the server does not charge for them.
What changed by default
Auto mode in Claude Code checks each action against a safety classifier before it runs. Until this release, the classifier ran as its own model call that was billed like any other call, so a session that made many small tool actions paid for a matching number of classifier requests.
Version 2.1.278 asks the server to run those checks as part of the same request that decides the action, and Anthropic says the server does not charge for that work. The default applies to Claude API accounts, Enterprise plans, Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry, and the Claude apps gateway. Pro, Max, and Team plans were never billed for classifier requests, so they see no change.
To check a live session, run /status. A new Auto mode server row reads Enabled when the server is running the checks and Disabled when the session has fallen back to Claude Code's own classifier calls.
When a session still pays
The server-side path only works if the request and reply reach Anthropic unchanged. An LLM gateway or proxy that strips headers, drops request fields it does not recognise, or edits streaming events can break it, and so can a platform, region or credential where Anthropic has not rolled the feature out yet. In either case Claude Code falls back to its own classifier requests, which are billed as before, and shows a one-time notice at the prompt naming the gateway when it can identify one.
The fallback is per session, not per action. Once the notice appears and the user presses Enter, Claude Code stops asking the server for checks for the rest of that session and does not show the notice again on the same machine for 24 hours. A single action the server could not check does not trip the notice; the tool handles that action locally and asks the server again on the next request.
On Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and signed-in Claude apps gateway sessions, only Claude Sonnet 5, Opus 4.7 or later, and the Fable models support auto mode at all.
The opt-out and the environment variable
Teams that run through a gateway they know cannot pass the request through can set CLAUDE_CODE_AUTO_MODE_SERVER=0 in the shell or in the env settings key before starting a session. Claude Code then never asks the server for checks, uses its own classifier calls, and does not show the notice. Anthropic notes in the documentation that this variable is temporary and may be removed in a later release. The variable is ignored on a direct connection to the Anthropic API.
For gateway operators, Anthropic's LLM gateway protocol reference lists what a gateway has to pass through unchanged for the feature to keep working, including the safeguards request field and the safeguard_results field on streaming events. A gateway that forwards headers and body fields verbatim, and does not rewrite tool-use IDs, keeps auto mode on the no-charge path.
Source
Source: Anthropic
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.