Server-Side Classifier in Claude Code Auto Mode: Billing, Gateways, and Fallback Logic

If your terminal pauses during command execution in Auto Mode after updating Claude Code and displays the following warning:

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

it means the client attempted to activate the new zero-cost server-side safety validation, but infrastructure constraints prevented it from applying to your active session.

In Auto Mode, a classifier examines potentially sensitive actions—such as Bash executions, system commands, and outbound network calls—before they run. Prior to version v2.1.278, these checks were performed via separate client-side requests that consumed billable tokens. Beginning with v2.1.278, validation is offloaded by default to Anthropic and cloud platform servers: checks now run directly within the primary model requests at no extra charge.

When server-side validation is unavailable, Claude Code continues verifying actions via client-side classifier requests. These fallback requests are billed at standard token rates as before; the warning simply indicates that validation has fallen back to billable client-side classifier requests, not that safety checks are disabled.

---

Responding to the Notice: Enter vs. Esc / Ctrl+C

When server-side checks are unavailable, Claude Code pauses execution before running the first command that requires validation and waits for your response:

• Pressing Enter: Approves the delayed command. The session continues in Auto Mode using client-side classifier checks, which are billed at standard token rates. If an intermediate gateway or proxy was explicitly identified in the prompt, this confirmation is cached on the local machine for 24 hours. If no gateway name was identified, the prompt will reappear upon the next session fallback.
• Pressing Esc or Ctrl+C: Aborts the pending action immediately and terminates the active turn. The session remains in Auto Mode, no confirmation is saved, and the prompt will reappear on the next command requiring validation.
• Switching modes: Press Shift+Tab to exit Auto Mode entirely. Disabling security policies or sandbox protections to avoid token charges is strongly discouraged.

In non-interactive environments, Claude Code adapts its behavior for automation:

• In headless mode (-p), the notice is directed to stderr while execution proceeds.
• In stream-json mode, the client emits a system warning event into the message stream (consumable via the Agent SDK).
• In the VS Code extension, the notice is displayed as an informational banner in the chat UI and requires no keyboard confirmation.

---

Session Diagnostics: The /status Command and Fallback Logic

To inspect how Auto Mode is currently handling safety checks, run the built-in status command:

/status

Look for the Auto mode server line in the output:

• Enabled: Safety validation is running on the server. No separate classifier token charges apply.
• Disabled: The session has fallen back to the client-side classifier. Safety requests are sent locally and billed as regular token consumption.

Isolated Failures vs. Session-Level Fallback

Distinguish between an isolated network hiccup and a persistent session fallback:

1. Single action: If the server fails to validate an isolated operation, Claude Code performs a single local check and attempts server-side validation again on the subsequent request. A session warning is not displayed in this case.
2. Session fallback: The on-screen prompt appears only when server-side checks fail for the entire session (which can occur as early as the first validated command).

---

Platform Availability, Regional Rollout, and Exceptions

Server-side validation is enabled by default in Claude Code v2.1.278 and later for:

• Claude API and Enterprise accounts;
• Claude Platform on AWS, Amazon Bedrock, Google Cloud Agent Platform (Vertex), and Microsoft Foundry;
• Compatible proxies and gateways.

Key availability rules and constraints:

• Pro, Max, and Team plans: Subscribers on these personal and team tiers never see this notice.
• Cloud model restrictions: On Amazon Bedrock, Google Cloud Vertex, and Microsoft Foundry, Auto Mode overall is supported only for Claude Sonnet 5, Opus 4.7 and newer, and Fable family models.
• Phased rollout: Server-side validation availability depends on platform, region, and credentials. If your network does not route through a proxy or corporate gateway yet /status shows Disabled, one possible reason is that validation is not yet available for this configuration. Check with your administrator or report via /feedback.

---

The Gateway Challenge: Transparent Pass-Through and the CLAUDE_CODE_AUTO_MODE_SERVER Flag

The most frequent trigger for this notice is an enterprise LLM gateway or proxy (such as a model router or load balancer) stripping or modifying traffic between the client and upstream API.

Gateway Requirements

Server validation requires transparent end-to-end metadata forwarding. To enable zero-cost server checks through a gateway, administrators must configure pass-through rules:

• Forward all request headers and body parameters unchanged, including security fields (specifically the safeguards field);
• Return responses and streaming events without stripping unfamiliar keys (specifically the safeguard_results payload);
• Preserve original tool execution identifiers (tool_use_id) without altering them at the proxy layer.

Temporary Opt-Out Flag and Trade-offs

If your enterprise gateway cannot pass these metadata fields and cannot be immediately updated, you can suppress the blocking prompt:

export CLAUDE_CODE_AUTO_MODE_SERVER=0

Key factors to understand about this environment variable:

• The trade-off: Setting this variable instructs Claude Code not to request server-side checks on Bedrock, Vertex, Foundry, or gateway routes. The warning prompt is suppressed, but all classifier checks will execute as billable client-side requests.
• Ignored on direct connections: This variable is not read and is completely ignored when connecting directly to the official Anthropic API.
• Setting CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 produces the same effect when CLAUDE_CODE_AUTO_MODE_SERVER is not defined.
• This setting is temporary and may be deprecated or removed in future Claude Code releases.

---

Cost Diagnostics: Event Timeline and Real Cost Drivers

A variety of assumptions have circulated regarding Auto Mode token consumption and HTTP 429 errors. Accurate cost assessment requires cross-referencing official release timelines and raw logs rather than hypotheses.

For official details on prompts, billing fallback, and gateway configuration, refer to the official Claude Code documentation.

Event Timeline

timeline
    title Evolution of the Auto Mode Verification Mechanism
    2026-09-08 : Historical incident #93558 : Specific 429 failure on versions 2.1.263–2.1.267 due to attribution header
    2026-09-19 : Official release v2.1.278 : Server classifier by default, indicator in /status, fallback warning
    2026-09 : Reddit discussion : Report of $50/hour spend in GitOps cycles without isolating causes

1. September 8–9, 2026 (Historical issue on legacy versions): Issue report #93558 documented failures on client versions 2.1.263–2.1.267. When combining a custom gateway with CLAUDE_CODE_ATTRIBUTION_HEADER=0, the client received a sequence of 15 empty HTTP 429 responses on classifier checks, blocking Bash commands. Removing the flag resolved the issue. The legacy client only restored attribution headers for the default api.anthropic.com host, dropping them on custom endpoints.
   • Note: This issue was specific to older client versions. The report contains no evidence that this recurs in v2.1.278, nor is every 429 status code caused by attribution header behavior.
2. September 19, 2026 (v2.1.278 release): Official release establishing server-side classification as default on supported platforms, adding the Auto mode server field to /status, and introducing the billed fallback notification dialog.
3. Reddit discussion on $50/hour spend: A user reported spending approximately $50 per hour across extended GitOps automation runs (Ansible, OpenTofu) on Opus 5-high. The user speculated that the Auto Mode classifier was transmitting full conversation history on every command. The thread provided no corroborating token accounting.

Why Classifier Costs Must Be Audited via Logs

Attributing a $50/hour bill directly to classifier overhead without log data is unsupported by evidence:

• While the Reddit poster captured JSONL session files, they did not perform a component-level cost breakdown to isolate classifier token usage from primary model inference. The actual expenditure drivers in that session remain unverified.
• Forum assertions alleging unannounced caching price adjustments by Anthropic represent user speculation rather than verified policy changes.
• Documentation confirms that server checks do not add extra billing, while client fallback requests are billed at standard token rates as before. A single user report is insufficient to quantify classifier overhead.
• In GitOps workflows, substantial command outputs (such as OpenTofu inventories/lists and voluminous Ansible output) are possible contributors to main-model context. This should be audited directly in session logs before drawing conclusions about classifier overhead.

---

Practical Troubleshooting Workflow for Engineers

If you encounter unexpected token usage or the fallback notice, follow this diagnostic checklist:

1. Verify your version: Ensure you are running Claude Code version 2.1.278 or later (claude --version).
2. Check your session state: Run /status and locate the Auto mode server line.
   • If Enabled: The classifier is running server-side without additional token charges.
   • If Disabled: The session has fallen back to billable client-side checks.
3. Isolate the root cause of fallback:
   • Behind an intermediate gateway: Check proxy logs to determine if safeguards or safeguard_results fields are being stripped, or if tool_use_id values are being modified. If the gateway cannot be updated immediately, use export CLAUDE_CODE_AUTO_MODE_SERVER=0 as a temporary workaround.
   • Direct API connection: Server validation may still be rolling out to your region or account tier. Confirm with your administrator or submit feedback via /feedback.
4. Audit expenditures using JSONL logs:
   • Open the session transcript (.jsonl) file;
   • Break down token counts (input_tokens, cache_read_input_tokens, output_tokens) by request type;
   • Separate safety classifier calls from primary model requests (Opus/Sonnet);
   • Measure the impact of verbose terminal output on the cumulative context window.
5. Maintain security baselines: Never disable safety guardrails or permission boundaries in an attempt to reduce token costs.