GlossaryFloor 1 · The Modela solid block on its own: the prediction machineFloor 1 · The Model
reasoning model
No. 059 · v2026-08FR: modèle de raisonnementA reasoning model first produces, for itself, a long series of intermediate steps, then writes its answer from them. Like someone covering a sheet with rough calculations before writing the line with the result.
What it is not
A reasoning model does not reason in the sense in which a person reasons: it produces a text of steps, and that text remains a prediction of what comes next, so it can be false while leading to the right answer, or the other way round. It is not an agent: it fetches nothing, calls no tool and does not act; all its activity lies in producing text before answering. And it is not a universally better model: on simple tasks it costs more, answers more slowly and brings nothing.
In depth
Moving the effort
The principle is to shift effort from training towards the moment of the answer. Rather than making the model bigger, it is left to produce many more fragments before concluding, and this behaviour is trained by reinforcement on tasks whose answer can be checked, mathematics and programming in particular. The gain is real and measurable on multi-step problems, where an immediate answer often goes wrong at the very first step.
Three costs
Three costs come with this gain, and they decide its use. Response time goes from a few seconds to sometimes several minutes. Price follows, since the intermediate steps are tokens produced, billed as such even when they are not shown. And variability rises: two runs may take different paths and reach neither the same result nor the same cost, which complicates budgeting as much as evaluation.
The standing of the intermediate text
The status of the intermediate text deserves to be understood. It is not a faithful explanation of the computation: it is a text produced by the same mechanism as the answer, and nothing guarantees that it describes what actually led to the result. You can therefore read a false step followed by a right conclusion. Taking it for an auditable justification would be a mistake, particularly in a field where a decision has to be reasoned. It helps spot a chain of reasoning going off the rails, it proves nothing.
Relations where the neighbours live
Check 3 questions · click your answer
Level 1 · Recognise
What does a reasoning model do that an ordinary model does not?
Level 2 · Distinguish
To rephrase a hundred product descriptions, is a reasoning model appropriate?
Level 2 · Distinguish
Can the text of steps produced by the model serve as the justification of a decision?
Lexigraph, "Reasoning model", v2026-08, https://www.lexigraph.org/en/reasoning-model/, CC BY 4.0.