Anthropic now charges for refused API requests

Anthropic's September 24 release notes add a line item API operators should read twice: the company is "resuming billing for refusals that arrive before any output" when the refusal's stop_details.category is "bio", "frontier_llm", or "reasoning_extraction". Those refusals are "charged like any other request, at the rates of the model that ran it."

This is the important part. A refusal is not an error. It arrives as a normal HTTP 200 with stop_reason: "refusal", an empty content array, and token counts in usage. You are now paying input-token price for a response that said nothing.

What bills and what does not

Anthropic's refusals-and-fallback doc lays out five classifier categories and which ones bill before any output. Billed: bio (requests that could enable biological harm), frontier_llm (requests that could assist development of competing models), and reasoning_extraction (requests that ask the model to reproduce its internal reasoning). Not billed: cyber, general_harms, and refusals with a null category.

Mid-stream refusals, the ones that cut off after partial output, were already billed. Nothing changes there. The billing applies on every platform Anthropic names: the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Billed or not, a refusal still counts against your rate limits.

Why the bio and frontier_llm categories matter most

Anthropic says it chose the three billed categories because they are "the categories where we measure low volumes of false positives, as of September 2026." That is the company's claim, not an independent audit. It also says the billed list "may change as Anthropic keeps measuring and refining its safeguards' false positive rates."

Here is the catch, from Anthropic's own table: "Beneficial life sciences work can also trigger this category" for bio, and "Benign machine learning work can also trigger this category" for frontier_llm. Those two categories now bill on arrival. If your pipeline runs Claude over genomics literature, protein data, or training workflows, you have false positive exposure that converts directly into invoice lines. There is no published false-positive rate per category and no estimate of what this adds to a monthly bill, so you cannot price the change in advance.

Anthropic frames the change as cost defense: billing exists "to disrupt attempts to circumvent Anthropic's safeguards at scale." Making automated jailbreaks and extraction probes expensive is a reasonable goal. The question operators get to ask is who pays for the misses.

The fallback double-bill

If you use fallback, the refusal that triggered it is billed in addition to the fallback request whenever the refusal arrived mid-stream or fell into a billed category. Fallback credit, which covers the prompt-cache miss on the retry, is unchanged.

This makes usage.iterations your ledger. It records every attempt, and Anthropic says tokens from different models are never summed into one field. If a refused turn fell back to another model, expect up to two billed attempts on that turn and reconcile them separately.

What to do this week

Log it before you get the bill. Every 200 response, not just errors, should capture stop_reason, stop_details.category, the model, token usage, and any fallback outcome. Split your refusal metrics into billable versus non-billable categories. For life-science and ML workloads, audit how often bio and frontier_llm fire on benign traffic; those are the false positives you will now pay for. If you run sub-agents that retry aggressively, cap the fan-out, because each refused attempt is now a chargeable event.

One more unknown worth noting: the docs describe how refusals are billed, but not how they will appear on an invoice. Per-attempt line items would make reconciliation straightforward. Aggregates would not. Watch the first bill after the change and check whether the categories you log match the charges you see.

These are API claims, documented but not independently tested. No keys were held and no requests were sent for this piece. The billing behavior above traces to Anthropic's own release notes and refusals documentation; interpretations of its cost exposure are ours.