Reasoning / Standard term
Large reasoning model (LRM)
A large reasoning model (LRM) is a model designed to spend extra internal computation when a problem benefits from deliberation.
A large reasoning model (LRM) allocates internal computation to problems that benefit from deliberation, such as constraint-heavy planning, proofs, and multi-file code changes. The term became common when providers offered separate models such as OpenAI o1 and DeepSeek-R1. Current products can expose reasoning as a setting or mode inside one model family. As of September 2026, Claude Fable 5.1 and GPT-6 Astra offer controls for how much reasoning they do. The choices available depend on the model and the app. A short classification can receive a direct response while a difficult migration plan receives more internal work from the same family.
Builder example
Model choice is now joined by effort and mode choice. A fixed rule that sends every difficult request to a separate reasoning model can waste time or miss tasks that need more exploration. Compare configurations on representative cases, including task success, required evidence, latency, and cost. Choose the configuration that clears the quality bar for that workload.
A delivery route must satisfy time windows, vehicle capacity, and driver hours at once.
Compare the current model's effort or mode settings on representative routes. Keep additional reasoning only when the same constraint checks show a useful gain.
Common confusion: Reasoning effort describes computation inside the model. Response length does not reveal how much reasoning occurred, and providers can return a short answer after substantial internal work. The label LRM describes a capability pattern; vendor product tiers change.