A reasoning model thinks in tokens before it answers in tokens
Ask a standard model a hard question and it starts writing the answer immediately. Its only "thinking" is whatever happens inside the network between reading your prompt and producing the first word.
A reasoning model works differently. It's trained to write out a chain of working first, often thousands of tokens of it, and only then write the answer. The working is text the model generates for itself. It tries an approach, notices a contradiction, backtracks, checks arithmetic, and eventually commits. Most products hide this chain or show a summary of it, and label the whole thing "thinking".
The thinking is the same next-token prediction that what is an LLM describes, pointed at the problem for longer. A reasoning model is a language model that has been trained, through reinforcement learning on problems with checkable answers, to spend more tokens before committing. That's why it helps most on problems where working can be checked and least on problems where there is nothing to work through.
Where the idea came from, in three dates
In January 2022, Jason Wei and colleagues at Google showed that putting a few worked, step-by-step examples in the prompt raised a model's scores on maths and logic problems. They called it chain-of-thought prompting, and it worked because writing the steps out gave the model more tokens, and therefore more computation, per problem.
In September 2024, OpenAI released o1, the first widely available model trained to do this by default rather than on request. The reinforcement learning behind it rewarded chains that led to correct answers on maths and code, so the model learned to produce long, self-correcting working on its own.
In January 2025, DeepSeek released R1 with open weights and a paper describing the training recipe, which meant every lab could reproduce the approach. Within the year, reasoning became a mode rather than a model: a toggle inside GPT, Claude and Gemini, and a family of dedicated rows such as Kimi K2 Thinking in the wider catalogue.
When thinking pays for itself, and when it is pure cost
The pattern across every benchmark and every real task is the same: reasoning helps in proportion to how many steps the problem has and how checkable the answer is.
| Task | Thinking mode | Why |
|---|---|---|
| Multi-step maths, word problems, unit conversions | Worth it | Steps can be checked; errors compound without working |
| Debugging, tracing a failure through code | Worth it | The model tests hypotheses against the code before committing |
| Planning a schedule, a migration, a budget with constraints | Worth it | Constraints interact; a first draft usually violates one |
| Logic puzzles, proofs, rules with exceptions | Worth it | The whole point is the chain |
| Factual lookup ("what year did X happen") | Wasted | No steps; the model knows or it does not |
| Rewriting, translating, summarising, changing tone | Wasted | One pass is the task; thinking adds delay and no quality |
| Casual conversation, brainstorming a list | Wasted | Nothing to check; more tokens for the same answer |
Thinking doesn't fix hallucination on facts. A model that doesn't know the capital of a country will reason its way to the wrong one with more confidence. Long chains also go wrong in ways a short answer can't. The model talks itself out of a correct first instinct. On easy questions, a long chain sometimes ends somewhere worse than the plain model's first answer.
What thinking costs: the tokens, the seconds and the credits
The thinking chain is output tokens, the expensive kind, and what are tokens in AI shows that output costs five to six times input on every major model. A reasoning turn that writes 3,000 tokens of working before a 300-token answer has generated ten times the output of the plain reply. It also takes roughly ten times as long, because output is produced one token at a time.
Leaderboards now price the dial. Artificial Analysis scores each model at every effort setting its provider exposes, so the tax is measurable: the same model, at its lowest and highest reasoning effort, with its Intelligence Index score above each end. The three ladders below were read on 2026-09-17.
Plot every setting and the curves cross. GPT-6 Astra reaches 53 at $2.31 per million tokens. Claude Fable 5.1 reaches the same 53 at $5.98. Claude Opus 5 at max effort lands on 51, two points below both, at $5.86. Effort buys points on every curve. Your curve decides what they cost.
Whizi's Auto router turns that into a fixed price you can see. Auto reads each message and sends it to one of six rungs; the rung is what sets the credit cost, and multi-step reasoning sits at the top on Powerhouse. Pro tops out at the 10-credit code rung, and Starter stays on the house model at 1.
A reasoning turn is 20 credits against 1 for a quick question: a 20-to-1 spread that mirrors the token cost underneath. Auto also refuses to waste it. A fresh message it classifies as a rewrite or a lookup goes to the 1- or 2-credit rung, and a follow-up in a thread that's already earned the reasoning rung stays there, because the previous turn's model acts as a floor. How Auto picks a model has the rung table with the model behind each row.
How to use thinking mode without paying for it everywhere
Treat thinking as a tool you reach for, not a setting you leave on.
- Default to the fast model. Most messages are lookups, rewrites and short questions. Thinking adds seconds and credits and changes nothing.
- Switch on for the hard turn. When the question has several steps and a checkable answer, move that one message to a reasoning model or toggle thinking, then move back.
- Give it the constraints in the prompt. A reasoning model is at its best when it has rules to check its plan against. "Must fit in 40 hours, no meetings on Fridays, the launch is the 12th" is the kind of prompt that earns the wait.
- Read the working when it is shown. The chain is where the model reveals a wrong assumption. If the first step is wrong, the answer is wrong however long it thought.
- Do not use it to check facts. Use search, a document, or a second model for that; what is an AI hallucination explains why more thinking does not help there.
If you'd want to see someone's working before trusting their answer, use a reasoning model. If you wouldn't, don't pay for it.
- Keep the fast model as the default and reach for thinking mode on the specific hard turn
- Use reasoning for maths, logic, planning with constraints and debugging
- Skip it for lookups, rewrites, translation and summaries; it is slower and no better
- Put the constraints in the prompt so the model has something to check its plan against
- Verify facts with search or a second model, not with a longer chain of thought
Frequently asked questions
What does a reasoning model actually do differently?
It writes out a chain of intermediate steps before the final answer, trying approaches, checking arithmetic and backtracking, all as generated text. A standard model starts writing the answer immediately. The reasoning is the same next-token prediction, spent for longer on the problem, which is why it helps on multi-step tasks and not on lookups.
Is thinking mode always better?
No. On factual lookups, rewrites, translation and casual questions it is slower, costs more tokens and doesn't improve the answer. On easy questions a long chain can talk the model out of a correct first instinct. It is better on maths, logic, planning under constraints and debugging, where the answer has steps that can be checked.
Why is thinking mode slower and more expensive?
Because the thinking is output tokens, produced one at a time, and output costs five to six times input on every major model. A turn that writes 3,000 tokens of working before a 300-token answer generates ten times the output of the plain reply and takes about ten times as long.
Which models are reasoning models?
OpenAI's o1 in September 2024 was the first widely available one, and DeepSeek R1 in January 2025 published the recipe with open weights. Since then reasoning is usually a mode inside GPT, Claude and Gemini rather than a separate product, alongside dedicated rows such as Kimi K2 Thinking. Whizi routes multi-step questions to its reasoning rung automatically.
Does a reasoning model hallucinate less?
On multi-step problems, yes, because it catches its own arithmetic and logic errors in the chain. On facts, no: a model that doesn't know something will reason toward a wrong answer with more confidence. For facts, grounding with search or a document and a second model's opinion work; a longer chain of thought doesn't.
