Anthropic: Messages API Returns usage.output_tokens_details.thinking_tokens — What It Means for Billing and Streaming

The Anthropic Messages API now explicitly reports internal thinking tokens through the output_tokens_details.thinking_tokens field inside the usage object. This change lets developers precisely track the cost of Extended Thinking calls — but with an important caveat that may surprise teams who have not read the documentation carefully.

What is usage.output_tokens_details.thinking_tokens and why does it matter?

The field appears in every API response that uses Extended Thinking and has the following structure:

{
  "usage": {
    "input_tokens": 1024,
    "output_tokens": 512,
    "output_tokens_details": {
      "thinking_tokens": 10000
    }
  }
}

The key distinction: Anthropic bills for all generated thinking tokens, not the summarized view returned to the user. If thinking_tokens: 10000 but the summary in the thinking block contains only 500 tokens, the billing is for 10,000. The difference between visible and billed tokens can be a factor of 20 or more, depending on task complexity and model.

How does thinking block summarization work?

Claude Sonnet 4.6, Opus 4.6, and newer models use summarization (summarized thinking) as the default display. Internal reasoning is processed by a special model that compresses conclusions into a shorter form — that summary appears in the thinking field of the block, while the full token count is recorded in thinking_tokens.

Summarization does not affect thinking quality — Claude sees the entire internal process; only the output representation is reduced for the user. The signature field in the thinking block stores an encrypted version of the full content for multi-turn conversations and must be forwarded unchanged when continuing the session.

Streaming behavior and latency

Developers who want to reduce latency to the first text token can use display: "omitted" in streaming mode:

• The server does not emit thinking_delta events
• Streaming of thinking content is skipped entirely
• Result: faster time to first text token
• Cost remains unchanged — thinking tokens are still billed

Claude Fable 5, Mythos 5, and Opus 4.8 use display: "omitted" as the default, while Sonnet 4.6 and Opus 4.6 use display: "summarized".

Budget limits and cache interaction

The budget_tokens parameter sets an upper limit on thinking tokens per request and must be smaller than max_tokens. With interleaved thinking, the budget is shared across all thinking blocks in a session. Important limitation: changing thinking parameters (enabling/disabling, changing the budget) invalidates cache breakpoints for messages — systems that aggressively use prompt caching must account for this cost.

Maximum output limits: Opus 4.8, 4.7, 4.6 and Mythos Preview support up to 128k output tokens; Sonnet 4.6 and Haiku 4.5 up to 64k. With the Batch API and the output-300k-2026-03-24 beta flag, Opus models can generate up to 300k tokens.