What extended thinking costs, and why the number you see is smaller

Reasoning models produce two kinds of output and you're billed for both. Anthropic states it directly: the tokens spent reasoning are billed as output tokens even when the thinking text isn't returned to you. The display setting changes what you see and nothing about what you owe, and the docs carry an explicit warning that the billed count doesn't match the visible one. So a response that looks like 40 tokens of answer can carry several hundred tokens of reasoning underneath it, and anything estimating cost from response length is wrong by whatever the model decided to think. There are two further consequences worth knowing. Thinking counts toward your max_tokens, so reasoning and answer compete for the same budget and a cap sized for a plain response gets truncated once thinking starts. And thinking blocks from earlier turns that stay in context bill again as input tokens on every subsequent request.

• Output token counts look higher than the responses you can read.
• You estimated cost from response length and the bill came in above it.
• Responses stop mid-answer with a max_tokens stop reason.
• Cost per call rose after enabling reasoning and nobody sized it first.
• Your cache hit rate fell after someone tuned the effort setting.

1. Read the thinking breakdown on a recent call and compare it to the total output.
2. Work out what share of your output tokens is reasoning rather than answer.
3. Check whether any responses are stopping at max_tokens.
4. See whether effort is set once per conversation or changing between turns.

Anthropic's thinking documentation, read 2026-08-05, demonstrates that changing the effort level invalidates the prompt cache, and publishes the usage from a three-request script on Claude Opus 4.8. Request two reads 3,546 cached tokens. Request three, identical except that effort moves from the default to medium, re-creates the same 3,546 rather than reading them. Rates are the price book's real Opus 4.8 figures: $0.50 per million for a cache read and $6.25 for a 5-minute cache write. Both the token count and the behaviour are published.

Twelve and a half times, on a change made to save money. The cache multipliers are published and the token count is Anthropic's own, so this isn't an edge case someone constructed. It's what happens whenever a conversation retunes its effort between turns, and it scales with how much context you had cached.

The reasoning share is reported per call and nowhere else. It isn't on an invoice, it isn't derivable from a response, and it varies request by request because the model decides how much to think based on what you asked. That makes it exactly the kind of number that has to be captured at the call or not at all. Culpa records the totals the provider reports and prices them from a versioned price book in exact decimal, so reasoning cost lands inside your per-call figures rather than as a surprise in the monthly total, and it carries whatever identifiers you attach. The question worth asking then stops being what thinking costs in general and becomes which feature, prompt version or customer is triggering it most, which is answerable from your own traffic and from nothing else.

• Tool Definition Token Cost: You Pay for Every Unused Tool
• Output Token Creep: The Rise Input Monitoring Can't See
• What is token accounting
• Image Token Cost: an Image Is Input Tokens, on a Formula