Key takeaways
- A reasoning model spends tokens on internal working before it answers; OpenAI and Anthropic both bill those tokens as output.
- The DeepSeek-R1 paper reports that reinforcement learning alone, without human-written reasoning examples, produced reasoning on checkable tasks like math and code.
- Both labs expose an effort setting, and Anthropic says Claude 4.7 and later reject the old fixed thinking budget.
A reasoning model is a language model that writes out a long stretch of working before it gives you an answer. Those working tokens are hidden or summarized, you pay for them, and they are why such a model is slower and costlier than a standard one on the same question. It is better at multi-step problems for the same reason: it gets room to plan, check and try again.
How a reasoning model works
An ordinary model reads your prompt and starts writing the answer. A reasoning model first spends tokens on internal reasoning, then writes the answer. OpenAI's guide to reasoning models describes this directly: the model uses reasoning tokens before it responds, which helps with planning, tool use and multi-step tasks.
Anthropic's docs say the same thing in its own terms. Claude can think before it answers, and its thinking is adaptive: on each request the model decides whether to think and how much, according to Anthropic's page on steering thinking. A simple factual question might get no thinking at all, while a tricky debugging task triggers deeper reasoning.
So "reasoning model" describes a behavior more than a separate kind of product. Some models always think, some choose, and most labs now let you set how hard.
How reasoning models are trained
The best-known account of how this was learned is the DeepSeek-R1 paper. Its authors report that reasoning in large language models can be encouraged through reinforcement learning alone, without human-labeled reasoning examples. Reinforcement learning means the model is rewarded for correct results and adjusts itself toward whatever earns the reward.
According to the paper's abstract, this led to patterns such as self-reflection, verification and changing strategy. It reports that on tasks with checkable answers, such as math, coding contests and STEM questions, the trained model beat counterparts trained on human demonstrations. The paper appeared in Nature, volume 645, in 2025.
The limit is in the words "checkable answers". Math and code can be graded automatically. Many everyday tasks cannot, and the abstract does not claim the method carries over to them.
What you pay for reasoning tokens
Reasoning is not free, and it does not show up as visible text.
- OpenAI: reasoning tokens are billed as output tokens. They are not visible through the API, but they take up space in the context window. The usage object reports them separately as
reasoning_tokens. - Anthropic: tokens Claude uses while thinking are billed as output tokens too. Billing covers the full thinking, not the summary you see, so the billed count will not match the visible text. The response reports the share spent on thinking in
thinking_tokens.
Both companies also warn about running out of room. OpenAI recommends reserving at least 25,000 tokens for reasoning and output when you start experimenting, because if the cap is hit, the response can end before any visible answer appears. Anthropic says thinking counts toward max_tokens, so a limit sized for a short reply is often too small.
How to control how much a model thinks
Both labs give you an effort setting instead of asking you to count tokens.
| Lab | Setting | Values listed in the docs |
|---|---|---|
| OpenAI | reasoning.effort |
none, minimal, low, medium, high, xhigh, max, depending on the model |
| Anthropic | output_config.effort |
low, medium, high, xhigh, max |
Lower effort means faster and cheaper answers. Higher effort means more thorough thinking. Levels vary by model: OpenAI says GPT-6 Astra rejects none, and GPT-6.1 Sol supports neither none nor minimal. Anthropic says Claude Opus 5.5 and Haiku 5.5 default to medium, while most others default to high.
Older Claude models used a fixed thinking budget instead. Anthropic now calls that legacy: Claude 4.7 and later models reject thinking.type: "enabled" with a 400 error.
When a reasoning model is worth it
OpenAI lists complex problem solving, coding, scientific reasoning and agent workflows. For work that gains nothing from reasoning, such as voice, quick lookups and classification, it suggests none effort, and low for tool use and planning when speed and cost matter.
A practical rule: start at low or medium effort, check whether answers are good enough, and raise it only for the tasks that fail. If you want to see how this ties into long agent runs, read what an AI agent is, or how prompt caching interacts with a changed effort setting. Anthropic notes that changing effort between requests invalidates cached prompts.
What we don't know
Neither lab publishes how many reasoning tokens a typical request uses, and the hidden reasoning cannot be audited by users. Treat the hidden reasoning as a cost to measure on your own workload, not as a guarantee of a better answer.
Frequently asked questions
What is the difference between a reasoning model and a normal LLM?
A reasoning model generates internal reasoning tokens before answering. A standard model starts the answer straight away. The extra working helps on multi-step tasks and adds cost and delay.
Are reasoning tokens billed?
Yes. OpenAI and Anthropic both bill them as output tokens, even though you do not see them in full.
Can I see a reasoning model's thinking?
OpenAI says reasoning tokens are not visible through the API. Anthropic returns a summary of the thinking, while billing for the full thinking.
How do I make a reasoning model think less?
Lower the effort setting: reasoning.effort at OpenAI or output_config.effort at Anthropic. Anthropic says this is more reliable than asking in the prompt.