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

Several bug reports have indicated that Anthropic's Claude Code is experiencing issues displaying summaries of its thought process, even though the underlying reasoning tokens still incur costs. When users request summarized thinking, the API has been returning empty thinking blocks for Opus 4.8 and Sonnet 5, despite the explicit request for summaries.

This issue, observed by developer Michael Hood as of July 16, 2026, seems to be a localized problem rather than a widespread concern. It might simply be an artifact of testing procedures that have changed how Anthropic displays summaries.

Another reported issue is the truncation of thinking block summaries, while the tokens billed are not adjusted accordingly. The company's documentation clarifies that users are charged for all thinking tokens generated, even if the summary is collapsed or redacted. This could lead to confusion about billing, as the company's terms state that customers are charged for all thinking tokens, regardless of whether they are returned in the output.

To mitigate costs, customers are advised to adjust their budget settings or disable thinking altogether.

Additionally, there have been reports of Claude Code terminating data streams during long-running thinking sessions. At least seven related API bug reports have been filed, with one identified by developer Hector Bernstorff as a client-side defect related to network tuning. Anthropic is actively working to address these issues, balancing latency concerns with the risk of requests getting stuck.

Community reports, such as GitHub issues, play a crucial role in quickly identifying and resolving problems. Anthropic appreciates the contributions of its developer community and is committed to addressing these issues as they arise.