Claude Code Auto Mode Billing: Server Classifier, Fallbacks and Costs

What changed in Claude Code v2.1.278?

On September 19, 2026, Anthropic released Claude Code version 2.1.278 with an important billing-related change to Auto Mode. Instead of issuing separate classifier requests by default, supported sessions can ask the server to perform safety classification within the session's model requests.

Anthropic's release notes say that the new server-side classifier is the default for Claude API and Enterprise users, and for supported Amazon Bedrock, Google Cloud Agent Platform, Microsoft Foundry and gateway configurations.

The release also introduces an Auto mode server row in the /status command so users can check whether the current session's classifier is running on the server.

The change does not eliminate charges for ordinary model work. It addresses the additional requests previously used to classify tool actions in eligible Auto Mode sessions.

The release notes also clarify that a session can fall back to the older billed classifier path and that Claude Code warns users when that happens on applicable accounts.

Official source: Anthropic Claude Code changelog, version 2.1.278.

How does Claude Code Auto Mode work?

Auto Mode is a permission mode that allows Claude Code to perform many actions without interrupting the user for routine approvals. It uses a separate classifier to review actions that require safety evaluation before they run.

The classifier checks whether a proposed action goes beyond the user's request, targets unrecognized infrastructure, risks exposing sensitive information or appears to be influenced by hostile instructions.

For example, an ordinary code-editing task may involve reading files, editing code and running a test suite. Auto Mode can let appropriate actions proceed without asking for confirmation at every step.

However, if Claude attempts an action that could irreversibly delete existing files, push sensitive data outside the repository or modify shared production infrastructure, the classifier may block it.

Claude can then try an alternative, request explicit human approval where appropriate or eventually fall back to manual approvals if repeated blocks prevent progress.

Source: Claude Code permission modes and Auto Mode safeguards.

What is the server-side Auto Mode classifier?

Before the new billing behavior, Claude Code could perform classification by making additional model requests on behalf of the session. Those requests carried some of the conversation context and the pending action, producing additional token consumption and latency.

In the new server-side path, the session asks the server to evaluate the action as part of its existing model-request flow. The resulting safety decision is used to allow or block that action.

Anthropic states that when the server performs the classifier review, the user is not separately charged for that classifier overhead.

The server-side change is primarily a change in where classification runs and how the extra safety-check work is charged. It is not a promise that every tool action will be approved or that Claude's main reasoning and coding work will be free.

When are Auto Mode classifier checks billed?

The central billing rule is straightforward: when the server performs the classification in an eligible session, there is no separate classifier charge.

If the server-side path is not available and Claude Code instead makes its own classifier requests, those requests may be billed under the applicable API, Enterprise or cloud-provider account rules.

Anthropic's documentation describes each separately billed check as including part of the conversation transcript and the pending action. This means the amount of classifier-related token usage can vary according to the task, conversation and number of actions needing review.

There is no universal flat classifier fee that applies to every Auto Mode session.

Does every tool call create a classifier charge?

No. Anthropic says ordinary reads and eligible working-directory edits can skip the classifier. The separate-request overhead primarily arises from actions such as shell commands and network operations that require review.

The total depends on which actions the session performs, whether the server-side classifier is active, and the applicable account's billing arrangement.

Official source: Claude Code documentation, Cost and latency and Server-side classifier review.

Which plans and platforms receive the classifier billing change?

Anthropic's August 2026 Auto Mode announcement had already removed classifier-overhead charges for Claude Code users on Pro, Max and Team plans.

The September 19 release extends the server-side default to API, Enterprise and supported cloud-provider or gateway configurations. Actual server-side eligibility can still depend on the platform, region, credentials, model and network path.

A listed provider is not a guarantee that every account, region, model and session is immediately eligible for server-side checks. Verify the current session instead of assuming the billing path from the provider name alone.

Official sources: Anthropic's August 2026 classifier billing announcement and Claude Code v2.1.278 release notes.

What do you still pay for when the classifier is free?

The phrase “free Auto Mode classifier” describes a specific category of overhead. It does not change the normal accounting for a coding session.

On a metered API account, your main model's ordinary input, output and applicable cache usage continue under the provider's rate card. On a Claude subscription, work continues to count against the plan's applicable usage limits, with any separately authorized extra usage handled under that plan's terms.

A task that runs for a long time may make many model requests, read and modify numerous files, launch supported subagents, or generate substantial output. Removing separate safety-classifier charges does not erase that underlying activity.

The correct question is therefore not “Is Auto Mode free?” but “Is this session receiving separate classifier charges in addition to its normal model and infrastructure costs?”

How to check whether your session uses the free server classifier

Claude Code v2.1.278 adds an Auto mode server row to the /status output. This helps distinguish a session using server-side classification from one using a different classifier path.

1. Check your Claude Code version. Open a terminal and run claude --version. The default server-side behavior documented here requires v2.1.278 or later.
2. Start an eligible Auto Mode session. Use your existing supported account, model and provider configuration.
3. Open the session status. Enter /status in Claude Code.
4. Look for Auto mode server. Review the displayed state and whether the session has encountered a fallback warning.
5. Check the state again after a fallback. Do not assume that a status observed before the session has exercised classification remains unchanged for the entire session.

claude --version

/status

A Disabled state is not, by itself, a universal billing verdict for every Claude account. Pro, Max and Team have their separate classifier-overhead policy, and a session not running Auto Mode is a different case from an Auto Mode session that has fallen back to billed classifier requests.

Official source: Claude Code v2.1.278 changelog and /status addition.

Why does Claude Code fall back to the billed classifier?

Anthropic explains that the server-side classifier must be able to review the pending action and return a usable verdict to the session.

If the server-side checks are unavailable or cannot reach the session properly, Claude Code can continue through its own classifier requests instead.

Common reasons include an incompatible LLM gateway or proxy, or a platform, region or credential that does not yet support the server-side path.

A gateway may accept ordinary model requests while failing to preserve the additional information required for server-side Auto Mode checks. That is why a normal Claude response does not prove that the free classifier path is working.

Once a session has entered the sustained fallback path, Claude Code can continue using its own classifier for the remainder of that session, with applicable classifier charges.

Official source: Claude Code documentation, Server-side classifier review.

What does the Auto Mode billed-fallback warning mean?

On accounts where separate classifier requests can be billed, Claude Code can display an eligibility notice when the server-side classifier path cannot be used for the session.

The warning explains that Auto Mode can continue through the older classifier-request path, but those additional checks may still incur token charges.

The important distinction is that the warning concerns the billing route for safety checks, not the availability of the core coding model.

If you accept the continuation option, expect the session to use the separately billed classifier path where applicable. If you do not want to proceed under that arrangement, cancel the pending action and investigate the server-side eligibility problem before continuing.

Does the warning mean the action itself is unsafe?

Not necessarily. A server-classifier eligibility warning and a classifier decision that blocks a risky action are different events.

The eligibility warning indicates that the intended server-side review path is unavailable for the session. An action block means that a classifier reviewed a proposed action and did not approve it.

Do not respond to an eligibility warning by disabling permission checks or using bypass-permissions mode merely to avoid additional classifier charges.

Why a gateway or proxy can prevent server-side classification

Some organizations send Claude Code requests through an LLM gateway or proxy rather than using a direct provider connection.

Gateways can help organizations manage routing, authentication, network restrictions and logging. However, they must handle the request and response information that Claude Code and the upstream provider need for supported features.

Anthropic identifies gateway or proxy interference as a common reason that server-side Auto Mode checks cannot reach a session.

Problems to investigate include:

• A gateway removing request headers required by the upstream service.
• A proxy dropping request fields it does not recognize.
• A gateway rewriting or omitting response fields associated with server-side review.
• Streaming-response transformations that prevent the client from receiving the expected results.
• A provider, region or credential where the server-side capability has not yet been enabled.

What should a gateway administrator do?

First, reproduce the problem in a controlled session and inspect the Auto mode server status. Confirm the Claude Code version, selected provider, model and relevant gateway configuration.

Next, consult the current Claude Code documentation for the server-side classifier requirements and compare the gateway's request and response handling against them.

Any compatibility change should preserve the required fields and identifiers without exposing credentials or bypassing organization-approved security controls.

After updating the gateway, begin a new eligible session and verify the status again. Do not assume that fixing the gateway automatically converts an already-fallen-back session back to server-side classification.

Official documentation: Claude Code server-side classifier review.

What does CLAUDE_CODE_AUTO_MODE_SERVER=0 do?

Anthropic documents CLAUDE_CODE_AUTO_MODE_SERVER=0 as an opt-out of the server-side classifier path on supported cloud-provider and gateway configurations.

Setting this variable tells Claude Code to skip requesting server-side classification and use its own classifier requests instead.

It is not a way to make classifier checks free. On accounts where the separate requests are billable, choosing this route preserves the older billed behavior.

export CLAUDE_CODE_AUTO_MODE_SERVER=0
claude

$env:CLAUDE_CODE_AUTO_MODE_SERVER = "0"
claude

These examples intentionally disable the server-side classifier path for the Claude Code process launched with that environment. They are not recommended as a general fix for unexpectedly high classifier costs.

Anthropic also says this variable is not read on a direct connection to the Anthropic API. Its effect depends on the supported provider or gateway setup.

If an organization uses the variable to maintain compatibility with an existing gateway, administrators should understand that it can keep the separate classifier-request billing behavior in place.

Official sources: Auto Mode server-side classifier settings and v2.1.278 opt-out behavior.

Can disabling experimental betas affect the server classifier?

Anthropic's permission-mode documentation identifies another configuration that can affect the server-side request path.

If CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 is set and CLAUDE_CODE_AUTO_MODE_SERVER is unset, Claude Code stops asking the server to perform those classifier checks.

This matters for organizations that disable experimental API features across their development environment. Such a setting may affect the classifier route even when the Claude Code installation is new enough to support server-side classification by default.

Before changing organization-wide configuration, confirm whether the variable is intentionally required for other compatibility or policy reasons.

Removing it without review could affect more than Auto Mode. Work with the administrator responsible for the environment rather than treating a billing issue as permission to change a managed configuration.

Official source: Claude Code documentation, Server-side classifier review.

Classifier-model fallback is different from billed server fallback

Claude Code has another fallback behavior that is easy to confuse with the September server-side billing change.

Anthropic documents Claude Sonnet 5 as the default classifier model, unless Anthropic configures another classifier model server-side.

The classifier model does not necessarily match the model selected through the session's /model command.

If the default classifier model is unavailable under the documented conditions, Claude Code can choose a supported fallback model for classification.

For example, Anthropic describes situations where the classifier uses the session's model or an applicable provider-default Opus model instead of Sonnet 5.

The first Auto Mode request validates the initial classifier choice. Once the validation settles, the selected classifier model remains fixed for that session.

Does classifier-model fallback automatically mean extra charges?

No. The classifier's model selection and the classifier's billing route are separate questions.

A server-side check can still have no separate classifier charge when the server performs it. A separately billed classifier request may use the model selected by Anthropic's fallback rules, subject to the applicable provider's billing arrangement.

Changing the selected coding model does not by itself tell you whether separate classifier charges apply. Inspect the active Auto Mode server path and the provider's actual usage records.

Official source: Claude Code documentation, Cost and latency.

When does Auto Mode fall back to manual approvals?

There is a third fallback mechanism that concerns user permission prompts rather than billing.

When a classifier blocks an action, Claude Code usually tries to find a different way to complete the task or asks the user for permission to proceed.

Anthropic's Auto Mode announcement states that Claude Code can fall back to manual approvals when it cannot make progress after three consecutive blocks or twenty blocks across a session.

This behavior is intended to prevent repeated automatic attempts from continuing indefinitely when the classifier refuses the proposed actions.

These mechanisms should not be treated as interchangeable. A billing fallback does not mean the classifier approved an unsafe action. A manual-approval fallback does not, by itself, explain a new classifier billing line item.

Official source: Anthropic's Auto Mode default announcement and manual fallback thresholds.

How to enable Auto Mode and verify that it is available

Auto Mode is a permission mode, so access depends on your account, model, organization settings, provider and Claude Code version.

Anthropic made it the built-in starting mode for eligible new Pro, Max and Team sessions in August 2026, subject to user and administrator settings. API and Enterprise sessions may continue to start in Manual mode unless another permitted default is configured.

The September server-classifier billing change does not mean that every existing Claude Code installation now starts in Auto Mode.

How to enter Auto Mode in the CLI

1. Update Claude Code to a current version and confirm your account and model support Auto Mode.
2. Start Claude Code in the project where you want to work.
3. Use Shift+Tab to cycle the available permission modes until Auto Mode is selected.
4. Alternatively, launch an eligible session with the Auto Mode permission flag.
5. Open /status to inspect the session's server-classifier state when you need to confirm the billing route.

claude --permission-mode auto

If Auto Mode is unavailable, check the selected model, provider, account eligibility and managed settings. Anthropic documents permissions.disableAutoMode as an administrative setting that can prevent Auto Mode from being selected.

If a session is not eligible for Auto Mode at all, it is not the same situation as an Auto Mode session that is active but has fallen back to separately billed classifier requests.

Official source: Choose a permission mode in Claude Code.

How to estimate the cost of Auto Mode classifier overhead

Anthropic does not publish one fixed classifier charge that applies equally to every Auto Mode task. Classifier usage varies with the actions that need review, the amount of context sent and the model and provider involved.

The simplest way to understand a metered account's total is to separate the cost of the coding task from any additional classifier usage.

Total relevant cost =
  Normal main-model usage
  + Separately billed classifier usage, if any
  + Other applicable provider or service charges

If the server performs all required classification under the no-separate-charge policy, the separately billed classifier component is zero for those server-performed checks.

If the session falls back to Claude Code-managed classification on a metered account, the additional component can be estimated from the billed classifier tokens and the applicable provider pricing.

For an API account, a general token-based estimate uses the applicable input, output and cache rates rather than assuming a fixed charge per action.

Classifier input cost =
  (Billed classifier input tokens / 1,000,000)
  x applicable input-token rate

Classifier output cost =
  (Billed classifier output tokens / 1,000,000)
  x applicable output-token rate

Add applicable cache charges or discounts
under the provider's rate card.

This is an explanation of the accounting method, not a claim that all providers expose identical classifier-usage fields or charge the same rates.

Why the same task can produce a different total

A task involving many reviewed shell commands can generate more separately billed classifier requests than a task that only reads files and makes eligible edits.

A longer conversation can also change the context sent to the classifier. Meanwhile, a session that uses server-side checks may eliminate the separate classifier-request component without reducing the underlying coding model's workload.

To investigate an unexpected bill, compare the model and token usage reported by the relevant provider with the session's classifier status and any billed-fallback warning.

Do not assume that upgrading Claude Code retroactively removes charges from earlier sessions or automatically refunds previous classifier usage. The release notes describe a new default billing path, not a blanket retroactive credit.

Common Auto Mode billing problems and how to check them

Most practical troubleshooting begins with identifying whether the problem concerns Auto Mode eligibility, the server-classifier route, the selected classifier model or ordinary model usage.

What if a gateway causes repeated fallback?

Gather the Claude Code version, provider configuration and the exact warning before changing settings. Confirm whether the gateway is intentionally modifying headers, request fields or streaming responses.

Work with the gateway administrator to restore compatibility. If the affected session has already entered a persistent fallback state, begin a new session after the configuration has been corrected and inspect /status again.

What if classification fails rather than falling back?

Anthropic documents separate classifier errors in which Auto Mode cannot determine the safety of an action. These can be transient, but some provider configurations require access to the classifier model selected for the session.

That is not necessarily the same as the billing-related server eligibility warning. Check the exact error and the model and provider requirements before applying a billing-related workaround.

Official guidance: Claude Code Auto Mode availability, server-side review and classifier cost behavior.

Practical checklist for avoiding unexpected classifier costs

1. Check the Claude Code version. Confirm that the installation supports the v2.1.278 server-side default.
2. Verify that Auto Mode is available and active. A Manual-mode session is not evidence of server-classifier billing behavior.
3. Inspect /status. Review the Auto mode server row in the active session.
4. Pay attention to fallback warnings. Do not assume that continuing under a billed classifier path has no cost.
5. Check gateway compatibility. If a proxy is involved, confirm it can preserve the server-side classifier request and response information.
6. Review relevant environment variables. Make sure the server classifier is not being intentionally disabled by a compatibility or organizational setting.
7. Separate classifier usage from main-model usage. Use the actual provider usage records when analyzing a bill.
8. Review high-impact actions. A free safety classifier is not a guarantee that an action is safe or appropriate for production infrastructure.