AI DevelopmentAnalysis5 min readPublished September 20, 2026

3 routers · 3 hiding places · vendor docs only · the decision before the answer

When Your Tool Picks the Model, Who Pays for Picking?

Claude Code 2.1.278 stopped charging for its auto-mode classifier where checks run server-side. What Claude Code, OpenRouter and Copilot bill for routing.

DA
Digital Applied Team
Research and practical guidance
Editorial dateSeptember 20, 2026
Claude Code 2.1.278September 19, 2026

Every auto mode, router or classifier makes a decision before the model you asked for does any work, and that decision runs on someone's compute. Until this week, most tools were quiet about whether you paid for it. On September 19, 2026 Claude Code 2.1.278 made the answer explicit for one product: where its auto-mode safety classifier now runs server-side, the classifier is not charged; where a session falls back to the old client-side requests, they are billed as before, and the tool says so.

That is the hook. The class is bigger: any team on a routed plan is paying for picking somewhere, and the question is whether the charge is documented and visible. This post defines the three places routing overhead hides, tables what three documented routers say they bill, gives you the one check to run on your own invoice, and lists what no vendor tells you. Numbers are quoted from vendor documentation; we have run no invoice arithmetic.

Key takeaways
  1. 01
    Claude Code now charges nothing for the classifier where the server runs it.Version 2.1.278 defaults to server-side checks on Enterprise plans, Claude API accounts, Bedrock, Google Cloud's Agent Platform and Microsoft Foundry. When checks cannot reach a session, usually because of a gateway, it falls back to billed classifier requests and shows a notice first.
  2. 02
    OpenRouter says the Auto Router adds no fee.You pay the standard rate of whichever model it selects. The catalog shows no price for the router entry, which means not quoted, not free.
  3. 03
    Copilot charges on the model auto picks, minus 10% on paid plans.GitHub's documentation says usage is charged on the selected model regardless of tier, with a 10% discount for paid plans, and that routing follows cache boundaries to avoid extra cache costs.
  4. 04
    A classifier charge and a model charge are different units.One is a safety check per action, the other is tokens per answer. Never add them into one number, and never compare a router's fee to a model's price as if they were the same thing.

01The hookWhat Claude Code changed on September 19

In auto mode, Claude Code runs a classifier over actions such as shell commands and network requests before they execute. The changelog for 2.1.278, published to npm at 01:48 UTC on September 19, says auto mode now defaults to a server-side classifier that does not charge for classifier overhead, for Claude API and Enterprise users and on Bedrock, Vertex, Foundry and gateways, with an environment variable to opt out on those platforms and a warning on billed fallback. The billing page explains the mechanism: the server performs the checks as part of the session's own model requests, so there is no separate request to bill.

The interesting part is the fallback. When the server's checks cannot reach a session, most often because a gateway or proxy strips a request field or drops a response key, Claude Code holds the next checked action and shows a notice saying the session is not eligible for the no-charge classifier. Press Enter and the session continues on billed classifier requests. Pro, Max and Team plans never see the notice. The status command gained a row that reads Enabled while server checks are deciding actions and Disabled once a session has fallen back. What the classifier actually decides, and how often, is in our post on Anthropic's published classifier numbers.

We're changing auto mode to no longer charge for classifier requests in Claude Code. However, this session isn't eligible.The notice text quoted in Claude Code's auto-mode classifier billing documentation, read September 22, 2026

02The classThree places overhead can hide

Routing overhead is any compute spent deciding rather than answering. It shows up on a bill in three shapes, and a vendor can document each one or not.

Shape 1
A classifier call
Per action or per prompt

A small model reads the prompt or the proposed action and returns a label: safe or not, which task type, which cost band. Billed as its own request, folded into the main request, or absorbed by the vendor. Claude Code's change is entirely about this shape.

Before the answer
Shape 2
A re-routed retry
Per failure

The primary model errors or is rate-limited and the router tries the next candidate. Whether the failed attempt is charged, and whether you can see it happened, is the question to ask.

After a failure
Shape 3
A plan multiplier or discount
Per plan

The router does not add a line item; instead the plan prices routed usage differently from fixed usage. Copilot's 10% discount for auto on paid plans is this shape in reverse.

On the plan

03The tableWhat each router says it charges

Three routers with published billing statements, read on September 22, 2026. Each cell paraphrases the vendor's own page; where a page is silent the cell says so. Cursor's router and other tools are not tabulated because we read no billing statement for them in this pass.

Claude Code auto-mode classifier billing page; OpenRouter Auto Router documentation; GitHub Copilot auto model selection, models and pricing, and legacy Copilot requests documentation. All read September 22, 2026; none prints a per-page date.
RouterIs the decision billed?Retries and fallbackVisible where?
Claude Code auto mode (2.1.278+)Not charged when the server performs the safety checks as part of the session's own requests. Billed as token usage when the session falls back to Claude Code's own classifier requests.Fallback is announced: a notice holds the first checked action and names a gateway when one is identified.The status command's Auto mode server row reads Enabled or Disabled; the fallback also prints to stderr under -p and emits a system warning in stream-json output.
OpenRouter Auto RouterNo additional fee: you pay the standard rate for whichever model is selected. The catalog prints no price for the router entry itself.The top-ranked models become the primary pick plus fallbacks; the documentation does not say a failed primary attempt is charged.The model field in each response names the model that answered; the task type is exposed only if you opt into router metadata.
GitHub Copilot auto model selectionUsage is charged on the model auto selects, with a 10% discount for users on paid plans in Copilot Chat, Copilot CLI, the GitHub Copilot app and Copilot cloud agent. Routing runs along cache boundaries to avoid extra cache cost.Not documented as a billable event; the reliability mode chooses by real-time system health, and retries are not described.Each response names the model that answered: hover in Copilot Chat, a terminal line in Copilot CLI, beside the model picker in the app. Billing units differ by plan: legacy annual Pro and Pro+ still count premium requests with per-model multipliers; every other plan bills AI credits.

Two of the three statements are worth reading in the original. OpenRouter's routing page says you pay the standard rate for whichever model is selected and that there is no additional fee for the Auto Router; its catalog lists the router with a price field of minus one, which is the catalog's way of declining to quote, and must not be read as zero. GitHub's auto model selection page says usage is still charged based on the model auto selects, regardless of tier, alongside a 10% discount for paid plans, and that switching models mid-session has shown increased cost without ample improvement, which is why routing happens at cache boundaries.

Do not add the columns

A classifier request is priced per check, a model answer per token, a Copilot premium request per prompt with a multiplier, a Copilot AI credit per unit of usage. They are four different units and this post keeps them apart. Any "routing costs X% extra" figure you see elsewhere has collapsed them into one and should state how.

04The checkThe one-line check on your own bill

Pick one routed session and count the model requests it made against the answers it returned. If the requests exceed the answers, the difference is overhead, and your vendor's documentation should tell you which of the three shapes it is. For Claude Code the check is even shorter: open the status command in an auto-mode session and read the Auto mode server row; if it says Disabled, the classifier requests on that session are billed and the notice should already have told you why.

How much overhead a harness adds on top of the model, before any router is involved, is measured in our harness-cost post; what a fixed monthly budget buys across tools is in our two-hundred-dollar comparison; and the per-token rates every router ultimately bills at are in the frontier model API price index.

05The gapsThe gaps nobody documents

Honest list. OpenRouter does not say whether a primary model that fails before returning is charged before the router falls back. GitHub does not say what the auto reliability mode costs when it steers away from a rate-limited model, or whether a routed prompt ever costs more than the same prompt on a fixed model. Anthropic does not publish the compute the server-side classifier consumes, only that it is not charged to you, and the opt-out variable is described as temporary. Cursor and other tools with routing modes publish no billing statement we could read for this post. If you run any of them at team scale, ask the vendor in writing which of the three shapes applies and where it appears on the invoice. Our AI transformation service includes that question in every cost review.

06Next stepThe decision before the answer is now a documented line for one vendor

Put it into practice

Ask each routed tool which of the three shapes it bills, and where you can see it

Update Claude Code past 2.1.278 if you are on the API, an Enterprise plan or a cloud platform, and check the status row on your next auto-mode session. For every other router your team uses, find the sentence in its documentation that says who pays for the routing decision. If there is no sentence, that absence is the finding.

Digital Applied

Know what your routed AI spend is actually buying.

We break a team's AI bill into answers, routing decisions, retries and plan multipliers, then set the modes and settings so overhead is visible and chosen.

Cost breakdownRouter settings reviewInvoice checks
Your next project

Start with one session

  • Requests made against answers returned
  • Which shape the overhead takes
  • Where it shows on the invoice
Questions and answers

Applying this post

Anthropic's documentation says those plans never show the eligibility notice. The server-side default and the no-charge statement are described for Enterprise plans, Claude API accounts and the cloud platforms; if you are on a subscription plan, nothing in the page changes your bill.
Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Continue reading

AI Development

Fable 5 Before July 7: The Six-Day Window Playbook

Claude Fable 5 was included through July 7 — now extended to July 12 — then moves to metered usage credits. How to sequence the window before the meter starts.

July 2, 2026 · 12 minRead
AI Development

When Checking AI Output Costs More Than Generating It

A cost formula for AI jobs at volume: generation, verification and rework. Worked at three failure rates and three checking designs on published 2026 rates.

September 20, 2026 · 6 minRead
AI Development

Fable 5 as the Planner Brain in Hermes and OpenClaw

Run Claude Fable 5 as the planner and route execution to cheaper models in Hermes and OpenClaw. The exact config keys, cost math, and marketplace hardening.

July 2, 2026 · 14 minRead
AI Development

Who Checks a Frontier AI Lab's Work? 12 Arrangements

A census of 12 external evaluation arrangements at Anthropic, OpenAI and Google DeepMind: who evaluates, who pays, what access they get, what gets published.

September 20, 2026 · 6 minRead
AI Development

A Proxy Stripped One Header and Claude Code Paid Twice

Claude Code v2.1.239 fixed what its changelog calls silently doubled billed API calls: behind a proxy stripping Content-Type, it re-ran turns non-streaming.

August 21, 2026 · 18 minRead
AI Development

Microsoft MAI-Cyber-1-Flash: Specialist Models Beat Big Ones

Microsoft's MAI-Cyber-1-Flash claims 96% on CyberGym at half the cost. The routing thesis behind it matters more than the unverified benchmark.

July 27, 2026 · 15 minRead