Anthropic’s Claude Code Displays Empty Thinking Summaries Amid Ongoing Reasoning Costs

Developers have reported that Anthropic’s Claude Code is returning empty or truncated thinking summaries, even when the feature is explicitly enabled. The issue affects Claude Opus 4.8 and Sonnet 5, with API responses delivering blank thinking blocks despite user requests for summarized reasoning.

Model “thinking” capabilities are designed to allocate additional tokens for complex problem-solving before generating a response, with summaries explaining the problem-solving process. While some platforms support this, Anthropic’s extended thinking functionality has been deprecated, and developers relying on transparency are now facing inconsistent output.

Software engineer Michael Hood observed that as of July 16, 2026, the API returns empty thinking blocks (thinking: ”, signature only) for Opus 4.8 and Sonnet 5, even when display settings are set to summarized, including when the request body is injected directly.

“As of 2026-07-16 ~15:00Z, the API returns empty thinking blocks (thinking: ”, signature only) for Claude Opus 4.8 and Sonnet 5, even when display: ‘summarized’ is explicitly requested — including when injected directly into the raw request body,” Hood noted.

Anthropic maintains that customers are charged for all thinking tokens generated, regardless of whether the text is returned or redacted. The company states that thinking tokens are billed as output tokens and count toward the maximum token limit, advising users to reduce budgets or disable thinking to control costs.

A separate report suggests that thinking summaries are being silently truncated while full token billing continues, with the anonymous author claiming a portion of the summary stream is dropped without adjustment. This has led to concerns that users are paying for reasoning output not actually delivered.

Beyond summary issues, the Anthropic API has been observed terminating data streams during extended thinking sessions. Developer Hector Bernstorff identified a client-side streaming defect related to network tuning, specifically involving request termination or retry mechanisms to balance perceived latency against the risk of stuck requests.

“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,” an Anthropic spokesperson told The Register. “We’re grateful to the developers who take the time to file them, and we’ll keep fixing things as they come up.”

Source link