What is the thinking token budget in Claude?

When extended thinking is enabled, Claude reasons internally before answering. The budget_tokens parameter caps how many tokens that internal reasoning can use. A larger budget lets the model think more on hard problems, at higher output-token cost.

How big should the budget be?

Anthropic suggests starting around 1,024 tokens minimum and increasing for harder tasks. Simple tasks rarely need more than a couple thousand; complex multi-step reasoning can justify 8,000 or more. This tool scales the recommendation by complexity and your preference.

Does the thinking budget count as output tokens?

Yes. Thinking tokens are billed as output tokens and count toward your max_tokens limit, so budget_tokens must be smaller than max_tokens with room left for the visible answer. The tool warns if the answer would not fit.

Will a bigger budget always give a better answer?

No. There are diminishing returns. Past the point where the model has fully reasoned through the problem, extra budget is wasted spend. Increase it only until quality stops improving on your evals.

Can I set the budget too low?

Yes. If the budget is too small for the task, Claude may truncate its reasoning and produce a weaker answer. Match the budget to genuine task difficulty rather than defaulting to the minimum on hard problems.

Token Budget Negotiator for Claude

Claude’s extended thinking lets the model reason before it answers, controlled by a budget_tokens parameter. Too low and hard problems get under-reasoned; too high and you pay for thinking you do not need. This tool recommends a budget that fits your task and your cost tolerance.

How it works

The recommendation combines three inputs:

Complexity baseline. Simple tasks start around 1,024 thinking tokens, medium around 4,000, hard around 10,000 — reflecting Anthropic’s guidance that harder reasoning warrants a larger budget.
Quality-vs-cost preference. A slider scales that baseline up (toward quality) or down (toward cost) by up to roughly 1.6×.
max_tokens safety. The recommended budget plus your expected answer length must fit inside max_tokens. The tool computes a suggested max_tokens and flags if the answer would be squeezed out.

It then estimates the thinking cost at your output price, since thinking tokens are billed as output tokens.

Worked example

A hard multi-step task, expecting a 1,500-token answer, slider at “balanced” (midpoint), output price $15/1M:

Baseline (hard): 10,000 thinking tokens
Balanced multiplier: ×1.15 → rounded to 11,520 recommended budget
Suggested max_tokens: 11,520 + 1,500 + 512 margin ≈ 13,500
Thinking cost per call: 11,520 × $15 / 1e6 ≈ $0.17

Slide all the way toward cost (×0.7) and the budget drops to about 6,912, cutting the thinking cost to roughly $0.10 — sensible if your evals show no quality loss.

Tips

Start from the recommendation, then tune on your own evals — raise the budget only while answer quality keeps improving.
Keep budget_tokens comfortably below max_tokens; the visible answer needs room.
Reserve large budgets for genuinely hard reasoning; simple extraction or formatting tasks barely benefit.
Model full spend with the LLM API Cost Calculator.