Claude Code’s Issues with Summaries
Anthropic’s Claude Code seems to be facing difficulties in presenting summaries of its “thought process,” as indicated by numerous bug reports, while the underlying reasoning tokens continue to incur charges. Users have reported that the API has been providing empty thinking blocks for Opus 4.8 and Sonnet 5 despite explicit requests for summarized thinking. Models from various providers can showcase their “thinking,” a mechanism that grants models extra tokens to navigate complex problems prior to generating a response. Software equipped with this feature provides a summary that clarifies its approach to a task. Some can also engage in “extended thinking,” although this feature is now obsolete. Developers frequently activate “thinking” hoping for improved outcomes, leading to additional token consumption and increased latency.
Observations from Developer Michael Hood
Developer Michael Hood has observed that certain models from Anthropic are not consistently effective at conveying their thought processes. “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 specifically requested — including when included directly in the raw request body,” Hood recently noted. It has been indicated that this issue is being looked into but does not seem to be a widespread, persistent problem. It may merely be an artifact of tests affecting how Anthropic showcases summaries.
Issues with Summaries and Billing
Similar instances of absent thinking blocks have been reported in Claude Code for VS Code. Another bug report suggests that thinking block summaries are being cut off while token charges remain unchanged. “The thinking is generated (and billed) in full; a section of the summary stream is silently omitted,” claims the anonymous author. This assertion, that clients are being charged for undelivered text, might stem from a misinterpretation of Anthropic’s policies: “You are billed for all thinking tokens generated, even when collapsed or redacted,” explains the company’s documentation. Within that legal framework, a thinking summary is priced the same as the complete output. It remains unclear if truncation due to bugs would alter the billing situation. “Thinking incurs a cost: the tokens Claude utilizes for reasoning are charged as output tokens, even if the thinking text isn’t provided to you, and they contribute to max_tokens along with the response text,” the company clarifies.
Concerns with Network Tuning and Terminations
Additionally, the Anthropic API has been observed to terminate data streams during protracted thinking sessions. At least seven other related API bug reports exist, but the streaming issue outlined by developer Hector Bernstorff highlights client-side faults. GadgetLad understands that this specific issue pertains to tuning network behavior, specifically aimed at terminating or retrying lengthy requests. Efforts are underway to balance perceived latency against the likelihood of requests becoming stuck.
Community Input and Continuous Resolutions
“Claude Code releases updates almost daily, and community contributions like these GitHub issues play a crucial role in how we rapidly identify problems,” an Anthropic representative informed GadgetLad. “We appreciate the developers who invest time in reporting them, and we’ll continue to resolve issues as they arise.”
Conclusion: The Token Dilemma: The More You Contemplate, the More You Spend
It seems that Claude Code is seriously contemplating how to express its thoughts, yet it’s left with more blank spots than a poorly written novel. And while you ponder this, those tokens are accumulating like unwashed dishes post-Sunday lunch. It’s high time Anthropic pulls itself together and addresses this situation before everyone runs out of funds!
