Reasoning & chain of thought
What are reasoning models?
Reasoning models are trained to do the step-by-step thinking on their own, producing an internal chain of thought before the final answer instead of waiting for you to ask.
Chain of thought started as something you had to ask for. Newer models build it in. A reasoning model is trained to think through a problem first, on its own, before it writes the answer you see.
OpenAI describes its reasoning models as using “internal reasoning tokens before producing a response,” which the model spends to “plan, use tools effectively, inspect alternatives, recover from ambiguity, and solve harder multi-step tasks.”[1] Anthropic’s version is called extended thinking: Claude works through a problem in its own words before answering, restating what is being asked, trying approaches, and checking intermediate results, and that reasoning arrives as thinking content blocks the final answer draws on.[2]
The practical difference is who does the prompting. With an ordinary model you add “think step by step” yourself. With a reasoning model the thinking happens automatically, and you often see it labeled separately from the final answer. That thinking is not free: it uses extra tokens and adds time, so these models cost more and respond more slowly than a plain question-and-answer model, anywhere from a few hundred to tens of thousands of reasoning tokens depending on how hard the problem is.[1]
What you actually see of that thinking is usually less than all of it. OpenAI is explicit that “reasoning tokens are not visible via the API,” even though “they still occupy space in the model’s context window and are billed as output tokens”; what you can see, if anything, is a summary generated separately from the raw reasoning.[1] Claude’s extended thinking works the same way in practice: a response can contain “summarized thinking blocks” rather than the complete internal trace, and a separate token count in the response tells you how many of the billed output tokens went to reasoning you never directly read.[2] You are paying for work you cannot fully inspect, on both platforms.
How much a model reasons is also not fixed. OpenAI’s API exposes a reasoning-effort setting running from none up through minimal, low, medium, high, xhigh, and max, trading speed and token cost against how completely the model works through a problem.[1] Claude’s manual mode instead uses a token budget: budget_tokens sets a target for how deeply the model thinks, actual usage varies and Claude may stop well before the budget is spent, and a separate max_tokens setting is the hard ceiling on total output.[2] Either way, the choice that used to be entirely yours, whether to ask for reasoning at all, becomes a dial on how much of it to pay for.
Reasoning models
- OpenAI reasoning models ↗
The o-series and GPT-5 reasoning models think internally before responding.
- Claude extended thinking ↗
Claude produces step-by-step thinking blocks before its final answer.
References
- Reasoning models — OpenAI
- Building with extended thinking — Anthropic