Reasoning models: what does the thinking cost?

Reasoning models write a hidden chain of thought before they answer. It makes them much better at math, code and planning, and you pay for every one of those hidden tokens at the output price. Here is what that means in numbers.

Thinking has taken over the top

70% of the text models we track can reason, and so can 20 of the 20 highest-rated models (10 of the top 10). Among rated models, the median reasoning model scores 148.4 on the overall index, against 127.2 for models that answer directly.

The price of a thought

Reasoning models also cost more per token: their median output price is $3.75 per million tokens, against $0.53 for the others. And the thinking itself is billed as output. With GPT-6 Astra, an answer that reads 1,000 tokens and writes 500 visible ones costs $0.035. Add 2,000 thinking tokens and it costs $0.135; with 10,000, an amount hard problems can reach, $0.535: 15× the price of the direct answer.

When thinking is worth it

• Worth it: math and logic, multi-step planning, code that has to work, analysis where a wrong answer is expensive.
• Not worth it: short replies, extraction, classification, translation, small talk. A model that answers directly is faster and cheaper there.

How to keep it under control

• Many models let you set how much they think (a "reasoning effort" or thinking-budget setting). Use the lowest level that still gives good results.
• Set a maximum output length, so that a runaway chain of thought cannot inflate a single answer.
• Route requests: send only the hard ones to a reasoning model. The model mix simulator shows the saving.
• Estimate before you ship: the cost calculator has a field for thinking tokens.

"Can reason" here means the model offers a thinking mode according to its listing; many models can switch it on or off per request, and you only pay for the thinking they actually do.

The figures in this article are recalculated from our data at every update (last: Sep 30, 2026). How we work