Developers are reporting cases where Anthropic’s Claude tools fail to show the “thinking” summaries they requested, according to The Register, even though the underlying reasoning tokens may still count toward the bill under Anthropic’s published terms. The issue centers on thinking blocks: summaries of the reasoning Claude uses before producing an answer. The Register says developer complaints describe empty thinking blocks for Claude Opus 4.8 and Sonnet 5, including cases where users explicitly requested summarized thinking. Developer Michael Hood reportedly traced the behavior to July 16, 2026, around 15:00 UTC, saying the API returned an empty thinking field with only a signature even when the request asked for summarized display. The same report says related behavior has appeared in Claude Code for Visual Studio Code. A separate bug report cited by The Register alleges that thinking summaries are being truncated while billing is not reduced, with the author claiming that thinking is generated and billed in full while part of the summary stream is silently dropped. That billing complaint is where the distinction matters. The Register notes that Anthropic’s own documentation says users are charged for thinking tokens generated, including when the thinking text is collapsed, redacted, or not returned. Anthropic’s docs also say those reasoning tokens are billed as output tokens and count toward max_tokens alongside the final response. In other words, the missing visible summary and the billing behavior are not necessarily the same bug. The display problem may prevent developers from seeing the reasoning summary they expected, but Anthropic’s documented billing model already allows charges for reasoning tokens even when the thinking text itself is not shown. The Register says it is unclear whether truncation caused by a bug would change that billing treatment. The report also says the issue is under investigation but does not appear to be a broad, ongoing concern. The Register was told the behavior may be tied to tests that change how Anthropic displays thinking summaries. Separately, the outlet says Anthropic’s API has been observed terminating streams during long-running thinking sessions, though one streaming issue it cites involves client-side defects and network tuning around whether to terminate or retry long requests. Anthropic told The Register that Claude Code ships updates nearly every day and that community reports, including GitHub issues, help it catch problems quickly. For customers trying to control spend, The Register points to Anthropic’s guidance: reduce the thinking budget setting or disable thinking. Who benefits: Developers who already meter token use closely are best positioned to manage this behavior by lowering thinking budgets or disabling thinking where it is not needed. Anthropic may also benefit if frequent Claude Code updates let it resolve display issues quickly. Who's exposed: Teams that enabled thinking primarily for transparency may be exposed if summaries disappear while reasoning costs remain. Cost-sensitive API users are also exposed if they assume visible thinking text maps directly to what is billed.