Claude API blank thinking blocks can still carry reasoning-token charges

Developers reporting Claude API blank thinking blocks may still be paying for the model’s reasoning tokens. Anthropic’s documentation says those tokens are billed as output whether the visible thinking text is returned or not, while complaints that requested summaries arrive blank point to a narrower, unconfirmed display problem rather than evidence that reasoning did not occur.

The distinction matters for teams using thinking-enabled models to handle coding and other multi-step work. A missing visible summary can be expected under one documented display setting. It is different from a response that returns an empty field after a developer explicitly asks for a summary.

Why does Claude return blank thinking blocks?

Anthropic says a thinking block contains a summary, not the model’s raw chain of thought. The block also includes a signature, described by the company as an encrypted copy of the full reasoning that must be passed back unchanged in multi-turn and tool-use conversations.

Under the documented display: "omitted" setting, the API returns a thinking block whose thinking field is empty. Anthropic says omitted display is the default on its newest listed models. Setting display: "summarized" is intended to return the reasoning summary instead. The company’s documentation gives thinking: {"type": "adaptive", "display": "summarized"} as the relevant configuration for models that accept it.

That setting changes visibility, not the token accounting. According to Anthropic’s thinking documentation, reasoning tokens count as output tokens and against the request’s max_tokens ceiling alongside the final answer. The company says the thinking block is billed the same whether its display is summarized or omitted.

What is the reported issue with requested summaries?

The Register reported that several bug reports alleged empty thinking blocks for Claude Opus 4.8 and Sonnet 5 even where summarized display was explicitly requested. Developer Michael Hood wrote that, as of July 16, the API was returning an empty thinking string and a signature for those models despite a summarized-display request.

The publication said it was told the matter was under investigation and did not appear to be a broad, ongoing problem. It also reported similar missing-thinking-block behavior in Claude Code for VS Code. That report does not establish that Claude Code’s display behavior has the same API billing treatment, or that every blank Claude Code block reflects generated reasoning.

A separate report alleged that summary streams could be cut short. The Register noted that it was unclear whether such truncation would affect billing. Anthropic’s published policy establishes charges for generated thinking that is hidden or collapsed, but the available evidence does not establish the cause, scale, or billing effect of the reports involving a requested summary that is blank or incomplete.

What can API teams control?

Teams that need to inspect the available summary should request summarized display on models and configurations that support it, then distinguish an omitted-display default from an unexpected empty response. For spend and capacity planning, Anthropic advises setting max_tokens high enough for reasoning plus answer text. Where a model permits it, thinking can be disabled; some listed models do not accept that option.

Anthropic told The Register that Claude Code updates nearly daily and that community bug reports help it find problems. No pricing figures, affected-request counts, or timeline for a fix were disclosed in the evidence reviewed.

This story draws on original reporting from The Register.