Anthropic expands which Claude API refusals get billed

Anthropic has widened the set of model refusals it charges for on the Claude API, adding three new refusal categories to the billed list starting today.

What's new

The change targets refusals that happen before the model produces any output. Anthropic's release notes describe it directly: "We're expanding which refusals are billed to include refusals that arrive before any output when stop_details.category is "bio", "frontier_llm", or "reasoning_extraction", the categories where we measure low volumes of false positives."

That last clause is the operative detail. Anthropic isn't billing every early refusal, only the categories where its own internal measurement shows the refusal is rarely a mistake: requests that trip bio-related safety filters, frontier-model extraction attempts, and reasoning-extraction attempts. Refusals that fall before output in other, less reliably-classified categories stay unbilled for now. Newly billed refusals in these three categories are charged at the model's standard per-token rate, the same as a normal completed request, and the change applies uniformly across the platforms Claude ships on, not just the direct API.

Context

How to charge for a refusal has been an open question across the API-model industry: a refusal still consumes compute to classify the request and decide to stop, but until now Anthropic treated most pre-output refusals as free. This is an incremental policy tightening rather than a new product, published as a dated entry in Anthropic's ongoing Claude Platform release notes, which track changes to the API, SDKs, and Console on a near-daily cadence.

Why it matters

For high-volume API users, especially those running agentic or bio/science-adjacent workloads where refusal rates run higher than average, this is a real cost input worth checking against current usage patterns, since it converts what was previously a free rejected call into a billed one whenever the stop category is bio, frontier_llm, or reasoning_extraction. Anthropic's own framing, that these are the categories with "low volumes of false positives," is also a signal about how confident the company is in its refusal classifier for those specific triggers versus the categories it left unbilled.