Reasoning Model
also called Thinking model, Hybrid reasoning
A model trained to spend tokens thinking before it answers.
A model trained — usually with reinforcement learning on problems that can be checked — to produce a long internal reasoning process before its final answer. You pay for the thinking in tokens and latency and get better performance on multi-step problems.
The practical difference from a standard model is that reasoning is a resource you buy rather than a prompting technique you apply. Where the API exposes a thinking budget, that budget is the main new dial: more thinking for hard problems, less for simple ones.
The token accounting is the practical change: thinking tokens are billed and they count toward the context window, so a reasoning model that "thinks" for two thousand tokens has spent real money and real budget before producing a visible word. Budgets and prompt design have to account for that.
The premium that bought nothing
A routing rule sends every request to a reasoning model after it wins a benchmark comparison.
On the extraction workload — which is most of the volume — latency triples and cost rises sharply with no accuracy change, because the task was never reasoning-bound. Restrict the reasoning model to the request class where it demonstrably wins on your eval set.
Using one for everything. On simple extraction and classification you pay a large latency and token premium for thinking that improves nothing.
- paperDeepSeek-R1: incentivizing reasoning capability in LLMs via reinforcement learning ↗
Reinforcement learning on verifiable tasks elicited reasoning patterns — self-checking, backtracking — with substantial open publication of the recipe.
- blogIntroducing Claude 4 ↗
Hybrid models with near-instant and extended-thinking modes, reported with coding and agentic evaluations — a concrete picture of what "reasoning model" means in a product.
- docsExtended thinking ↗
Configurable thinking budgets and interleaved reasoning around tool results — the API-level view of chain of thought as a resource you pay for.
Test-time Compute · Inference-time compute
Buying accuracy with more thinking at answer time, not more training.
Chain of Thought · CoT
Let the model work through steps before committing to an answer.
Inference
Using a trained model — the part you pay for on every request.
Model Routing
Sending each request to the cheapest model that can handle it.
LLM · Large Language Model
A very large model whose only job, at heart, is predicting the next token.