Ecosystem

Anthropic now bills refusals, even when nothing came out

2 min read AI-generated

Mid-stream refusals were billed already. What is new is the case where the answer never starts.

Featured image for "Anthropic now bills refusals, even when nothing came out"

A line in the Claude Platform release notes dated September 24 changes invoices. Anthropic is resuming billing for refusals that arrive before the model has produced any output at all. It applies to three values of stop_details.category: bio, frontier_llm and reasoning_extraction.

The reasoning is a measurement. In those three categories Anthropic sees low volumes of false positives, meaning few cases where the model treats a harmless request as dangerous. On that reading, a request refused there was mostly refused correctly.

Mid-stream refusals were billed already. What is new is the case where the answer never starts. Those are now charged like any other request, at the rates of the model that ran. For every other category nothing changes: if the reply dies before the first token, it costs nothing. Fallback credit is unchanged too. The switch applies on all platforms, not only the Claude API.

Two Compliance API changes on the same day

The local session endpoints are out of beta for Claude for Microsoft 365 — Excel, PowerPoint, Word and Outlook, recognizable by product_surface values beginning with office_agents.

The second change may break existing code. The Activity Feed no longer returns file names, project document names or artifact titles. The filename and title fields are now always empty or omitted, including on activities recorded before the change. If you need a name, look it up by the ID, which takes a Compliance Access Key with the read:compliance_user_data scope.

The number missing from the change is the interesting one

Anthropic justifies the billing with low false-positive volumes in three categories and gives no figure. For a change that costs money, that is thin. In practice it means this: if your work sits near biological topics, life-science research for instance, you now pay for the requests the safety mechanism throws away. Those same users are the ones most often refused by mistake. A cost item that used to be a rounding error becomes one worth watching — easiest through stop_details.category in your own logs.


Sources:

Claude APIPricingDevelopment