Claude extended thinking: when it earns its tokens
Eleven of the twelve Claude ids support thinking. A budget is spent on generated tokens and capped by the model output ceiling, so the output rate is the number that decides it.
Which ids support it, and what bounds the budget
Eleven of the twelve Claude ids in this catalog list thinking among their capabilities. The exception is stable-claude-haiku-4-5, which lists text and tools only — a request that asks it to think does not have a thinking mode to fall back on.
A thinking budget buys generated tokens, so it is bounded by the same ceiling as the answer: 128,000 tokens on the 1M-context ids (claude-opus-5, claude-opus-4-8, claude-opus-4-6, claude-sonnet-5, their stable-* twins, and both Fable ids) and 64,000 on claude-sonnet-4-6 and stable-claude-sonnet-4-6. A budget larger than the ceiling is clamped to it.
The budget also has to leave room for the answer itself. Setting budget_tokens close to max_tokens leaves the model nothing to reply with, so the useful pattern is to raise max_tokens along with the budget rather than spending the existing allowance on reasoning.
curl -sS https://token-share.app/v1/messages \
-H "x-api-key: $TOKEN_SHARE_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "content-type: application/json" \
-d '{
"model": "claude-sonnet-5",
"max_tokens": 8000,
"thinking": {"type": "enabled", "budget_tokens": 4000},
"messages": [{"role": "user", "content": "Two services deadlock under load. Here are both stack traces."}]
}'What the tokens cost
Thinking is spent in generated tokens, so it is priced at the output rate. On claude-sonnet-5 that is $15 per million official at the 0.1× multiplier — $1.50 as billed. On claude-opus-5 it is $25 official, $2.50 billed. On the Fable lanes it is $50 official at 0.4×, which is $20.00 per million.
Put a number on a single call. A 4,000-token budget on Sonnet 5 costs $0.006 of thinking; the same budget on stable-claude-fable-5-1 costs $0.08, because the output rate there is 13.3× Sonnet 5’s. Spending 32,000 tokens of thinking on Opus 5 is $0.08 per call.
At the level of one request these are small numbers. At the level of an agent that thinks on every step of a thousand-step run, they are the bill. Which is why the decision is worth making per task type rather than globally.
Where it changes the answer and where it does not
Thinking is reasoning done before the answer is committed, so it can only help where a first attempt would have gone wrong in a way that further work would catch. Problems with a long chain of dependent steps fit that shape: a mistake at step two is invisible in the output but wrong by step nine.
Work with no chain to get wrong does not fit it. Reformatting, extraction, classification, translation, a lookup against material already in the prompt — the answer is determined by the input, and more deliberation produces the same result at a higher output count.
The tell is whether your failures are wrong-answer failures or wrong-approach failures. A model that misreads a field will misread it after thinking too. A model that picks the wrong strategy and then executes it flawlessly is the case thinking is for.
How to measure it instead of guessing
Run your evaluation set three times on one id: thinking off, a small budget, and a large one. Hold everything else fixed. You get a curve of accuracy against output tokens, and the curve is specific to your tasks in a way no general claim about the feature can be.
Read two things off it. First, where the accuracy stops improving — that is your budget, and it is usually well below the ceiling. Second, whether the improvement is worth the output tokens it consumed, priced at the rate for the id you are running.
Then check the cheaper comparison. A larger budget on claude-sonnet-5 and a smaller one on claude-opus-5 can land at similar cost, since Opus output is 1.67× Sonnet output. Which of those two wins on your work is an empirical question, and the same three runs answer it.