Reasoning (LLM) — AI Glossary

Reference details

Related terms

• Chain of Thought (CoT)
• Tree of Thoughts (ToT)
• MATH Benchmark

How it works

1. A reasoning model is produced by reinforcement learning on problems whose answers can be checked automatically, so the reward is the final answer's correctness rather than a human judgement of the working.
2. Nothing supervises the trace itself, which is why the traces that emerge are not tidy explanations — they contain abandoned approaches, because abandoning one is what led to the reward.
3. That training signal is only available where answers are verifiable, which explains the shape of the capability: strongest on mathematics, code and logic, weaker on tasks with no checker.
4. Trace length is unconstrained at inference, so the model allocates more tokens to harder problems without being told to, and that allocation is itself learned behaviour.
5. The trace occupies the context window and is billed as output while usually being withheld from the caller, so cost and latency track a quantity the response does not reveal.
6. Because the reasoning was trained rather than prompted, instructing such a model to think step by step competes with its own learned procedure instead of invoking it.

Example: Reasoning that costs tokens you cannot see

A reasoning model may generate thousands of internal tokens before its first visible word. You are billed for them and you wait for them, but they never appear in the response. A task that looks like a 200-token answer can be a 4,000-token call — which is why reasoning models change latency and cost profiles, not just quality.

Commonly confused with

Reasoning models and chain-of-thought prompting are not the same thing. Chain-of-thought asks any model to show its steps in the output; a reasoning model was trained to reason internally before answering. Prompting the second for step-by-step output can fight the training rather than help it.

When to use it

Reach for it when:

• The task has verifiable intermediate structure: maths, logic, multi-constraint planning
• Accuracy matters more than latency or cost
• A wrong answer is expensive enough to justify the extra tokens

Reach for something else when:

• Retrieval, extraction, formatting or classification — reasoning adds cost, not accuracy
• The interface is latency-sensitive and users watch a blank space
• You are budgeting on visible output alone, which understates 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 (LLM) — AI Glossary.

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