Claude Code returns blank thinking blocks, but reasoning still costs you

Anthropic's Claude Code is reportedly experiencing issues with displaying summaries of its internal reasoning, despite still costing users money for the tokens used in reasoning. Users of the API have been receiving empty thinking blocks even when requesting summarized thinking, with the problem observed in Claude Opus 4.8 and Sonnet 5 as of July 16, 2026.

The API appears to be returning only the signature, not the actual reasoning summary. This issue may be a test artifact, but Anthropic is investigating it. Some developers have also reported truncated thinking blocks while being billed for the full output, which could be due to misunderstanding Anthropic's billing policy. The company charges for all generated thinking tokens, even when collapsed or redacted, and customers can reduce costs by lowering their budget or disabling thinking.

Additionally, the API has been seen terminating data streams during long-running thinking sessions, likely due to network tuning issues. Anthropic encourages community reporting of such bugs, as they help quickly identify and resolve problems.