Large reasoning model (LRM)
- Reasoning
- Standard term
A large reasoning model (LRM) is an AI model built to write out long step-by-step reasoning before answering, which helps with tasks such as planning under many constraints, proofs, and multi-step code changes.
Training, often with rewards for reaching answers a program can check, teaches the model to use that reasoning to try an approach, test it, and back up when it fails. The extra text makes answers slower and counts against your plan's limits or your bill, which is why apps let you raise or lower it.
Where you’ll meet it
The assistant built into your work spreadsheet drafts next month's staff schedule. It has to give everyone two days off in a row, put two people on every shift, and keep one person off Fridays. On its quick setting, it books the same person on two shifts at once.
What to do
Switch the assistant to its thinking or extended reasoning setting for this job and run it again. Then check the new schedule against each rule yourself before anyone sees it, because longer reasoning makes a clash less likely without ruling one out.
Easy to mix up
Prompted chain-of-thought writes the steps into the reply, where you can read them. A reasoning model writes long reasoning of its own, and the provider often shows you only a summary.
Check yourself
Your reasoning model is set to its highest effort, and the next job is converting a spreadsheet's dates into one format. Should you keep the setting there for this job?
Go deeperAccuracy on hard problems rose with longer reasoning, billed by the token
OpenAI's o1, released in 2024, made the category widely known. DeepSeek-R1 followed in early 2025 with a paper reporting that a version trained by reinforcement learning alone, given no worked examples to imitate, still learned to recheck and revise its own steps.
Providers usually bill the reasoning text as output, counted in tokens, which are pieces of words. On a hard question it can run to many times the tokens of the visible answer. Controls such as an effort level or a thinking budget may differ between a provider's API, which developers call from code, and its desktop app.
Researchers call this extra computation at answer time test-time compute. OpenAI reported with o1 that accuracy on hard math problems kept rising the longer the model was allowed to reason.
Sources