Anthropic now bills refusals that produce no output

Anthropic resumed billing for refusals on the Claude API — specifically, refusals that arrive before the model writes a single output token. Three categories are affected. If you run agents at volume, your refusal rate just stopped being free, and most teams have no idea what theirs is.

What actually happened

Per the Claude Platform release notes, dated September 24, Anthropic is billing pre-output refusals where stop_details.category is "bio", "frontier_llm", or "reasoning_extraction". Anthropic's stated reason: these are the categories where it measures low volumes of false positives.

The rest of the change is narrow, and the narrowness is the point:

• Mid-stream refusals were already billed. Nothing new there.
• Billed refusals are charged like any other request, at the rates of the model that ran it. There is no discount tier for a refusal.
• Refusals before any output in every other category are still not billed.
• Fallback credit is unchanged.

So this is not a price increase in the usual sense. It is a category of request that used to cost zero and now costs whatever your model costs — input tokens, thinking tokens, the whole bill — with nothing in the response body to show for it.

Why AI refusal billing matters for your business

The number you need is one you probably do not have: what percentage of your API calls terminate in a refusal, broken down by stop_details.category. For a chat product with human users, it is noise. For an automated pipeline hammering the same prompt shape ten thousand times a day, a systematic refusal is not noise — it is a stuck loop that now meters.

The nasty version is a retry wrapper. A lot of agent code treats a non-answer as a transient failure and retries with backoff. Under the old behavior, a pre-output refusal that your retry logic hit three times cost nothing three times. Now it costs three full requests, and it will keep costing them every run until somebody reads the logs. reasoning_extraction is the one to watch here, because it can fire on prompt shapes that look perfectly reasonable from the outside.

Two things to do this week. First, log stop_details.category on every response — not just the failures, all of them — and put a count on a dashboard. If you are already emitting OpenTelemetry from your agent harness, this is one attribute. Second, make your retry policy category-aware: a refusal is a deterministic outcome for that input, and retrying an identical prompt is buying the same answer twice.

This is the broader pattern with metered AI. The economics of your automation live in the edge cases, not the happy path, and vendors change the edge cases in changelog entries. Read the release notes for every model you depend on, or have something read them for you.

Sources: Claude Platform API release notes.