Anthropic's Claude Code is returning empty or truncated "thinking" blocks to developers while still charging them for the underlying reasoning tokens, according to multiple bug reports examined in a recent article by The Register. The technical issue affects Claude Opus 4.8 and Sonnet 5 models, with developers complaining that the API fails to deliver summaries of the model's internal reasoning process even when explicitly requested. The situation has raised questions about billing practices when AI systems generate—but don't display—the computational work users pay for.
Developer Michael Hood documented that starting around July 16, 2026, at approximately 3:00 p.m. UTC, the API began delivering blank thinking blocks for Claude Opus 4.8 and Sonnet 5, showing only the signature with no content, even when users set the display parameter to "summarized" and injected it directly into raw request bodies. The Register was informed the issue is under investigation but doesn't represent a widespread, ongoing problem, potentially stemming from experimental changes to how Anthropic presents summaries. Comparable problems with missing thinking blocks have surfaced in Claude Code for VS Code, while a separate bug report alleges that thinking block summaries are being cut off mid-stream without corresponding adjustments to token charges, with one anonymous complainant asserting that "the thinking is generated (and billed) in full; a portion of the summary stream is silently dropped."
According to Anthropic's documentation, customers are "charged for all thinking tokens generated, even when collapsed or redacted," meaning a condensed thinking summary carries the same price tag as complete output. The company explains that reasoning tokens Claude uses are billed as output tokens regardless of whether the thinking text reaches the user, and they count toward the max_tokens limit alongside regular response text. It remains unclear whether bug-driven truncation would alter billing calculations. The Anthropic API has also been observed cutting off data streams during extended thinking sessions, with at least seven related API bug reports filed, though developer Hector Bernstorff's streaming issue appears linked to client-side problems rather than server-side failures.
The core technical challenge involves tuning network behavior to decide when to end or retry long-running requests, with ongoing work attempting to strike a balance between managing perceived delay and preventing requests from getting stuck indefinitely. Models capable of displaying "thinking"—a feature granting additional tokens for reasoning through complicated problems before generating responses—attract developers who believe the extra computational steps yield superior results, despite higher token costs and increased latency. To cut spending on thinking functionality, Anthropic advises customers to reduce their budget settings or turn off thinking entirely. An Anthropic spokesperson told The Register that "Claude Code ships updates nearly every day, and reports from the community like these GitHub issues are a big part of how we catch problems quickly," adding gratitude to developers who file reports and a commitment to continue addressing issues as they emerge. The billing structure creates a transparency challenge that enterprise customers navigating AI budgets will likely scrutinize more carefully as reasoning-heavy models become standard infrastructure.