Anthropic: Advisor tool gains max_tokens, refused requests no longer billed

In its release notes for June 2, 2026, Anthropic published two changes to the Claude API aimed at better cost and latency control for development teams.

What does the new max_tokens for the advisor tool bring?

The advisor tool, a mechanism in which a separate advisor model provides guidance during tool use, now supports the max_tokens parameter. By setting tools[].max_tokens in the advisor tool’s definition, you can cap the length of the advisor model’s output per individual call. This is useful for workflows that do not need full-length responses, because it reduces both latency and the cost of output tokens.

What changes in the billing of refused requests?

The second change concerns billing. On the Claude API, a request is no longer billed when it returns stop_reason: "refusal" and Claude has not generated any output. In other words, if the model refuses a request for safety reasons without any text, that call does not enter the bill. To detect and handle such refusals, Anthropic points to the documentation on streaming refusals.

Why do these changes matter?

Both changes are small but practical optimizations for those building applications on top of the Claude API. Capping the advisor model’s output provides finer control over costs in agentic flows, while removing billing for refused requests eliminates the cost of calls that produce no useful result anyway.