Reasoning Model — AI Glossary

Reference details

Related terms

• Chain of Thought (CoT)
• RLHF (Reinforcement Learning from Human Feedback)
• Large Language Model

How it works

1. Training rewards the final answer on problems with checkable solutions, so the intermediate steps are shaped by whether they led somewhere rather than by imitating human explanations.
2. The model consequently generates a long internal chain before answering, and that chain is billed as output tokens whether or not it is shown.
3. Many providers hide the chain and return a summary, so the tokens paid for exceed the tokens displayed.
4. A compute knob — effort or thinking budget — trades accuracy against latency and cost at request time, which is a dial no prompting technique provides.
5. Longer chains help where a wrong early step can be detected and revised; on retrieval or formatting tasks they add cost and raise the chance of overthinking a settled answer.

Example: The hidden tokens are the cost

A request returns a 300-token answer after 3,000 tokens of internal reasoning. At a notional $15 per million output tokens the visible answer suggests 300 x 15/1,000,000 = $0.0045, while the bill is 3,300 x 15/1,000,000 = $0.0495 — about eleven times more. Nothing is wrong; the reasoning is the product. But a cost model built from visible output underestimates by roughly an order of magnitude, which is the budgeting mistake these models invite.

Commonly confused with

A model class, not a mode or a prompting style. Extended thinking is a mode that can be switched on, including on models not trained this way; chain-of-thought is a prompting technique that asks any model to show its working. A reasoning model was trained — typically with reinforcement learning against verifiable answers — so the deliberation is learned behaviour rather than something the prompt requests.

When to use it

Reach for it when:

• Problems with many dependent steps where an early error invalidates the rest — maths, proofs, intricate code.
• Agentic planning, where the quality of the plan dominates the cost of producing it.
• Accuracy is worth a multiple of the latency and token spend.

Reach for something else when:

• Extraction, classification, formatting or retrieval, where the answer needs no deliberation and the chain is pure cost.
• Interactive paths with a latency budget measured in hundreds of milliseconds.
• High-volume simple calls, where the hidden tokens dominate the bill.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Reasoning Model — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path