Developers have reported that Anthropic’s Claude Code API sometimes returns empty or truncated summaries of its model “thinking”, despite users requesting them. This matters because the underlying reasoning tokens may still be generated and charged for, leaving customers with higher costs and less visibility into how the model approached a task.

Reports affect Claude Opus 4.8 and Sonnet 5, with one issue said to have begun on 16 July 2026; similar behaviour has also appeared in Claude Code for VS Code. Anthropic says it charges for all reasoning tokens, including those collapsed or redacted, and advises customers to reduce the thinking budget or disable it to lower costs; it says the reports are under investigation and notes near-daily Claude Code updates.