A reasoning model is a large language model trained, typically with Reinforcement learning, to generate an extended internal sequence of reasoning steps before producing a final answer, rather than answering immediately as earlier-generation models did. The category is distinguished from ordinary Chain-of-thought prompting, which asks a conventionally trained model to reason step by step through instructions in the prompt, because a reasoning model is trained specifically to produce and improve its own reasoning process, and often generates far more intermediate text than a prompted model would.
Origins
OpenAI introduced the first widely known reasoning model, OpenAI o1, in September 2024, describing it as trained to "think" before responding using reinforcement learning on problems with verifiable answers, such as mathematics and programming, where a correct final answer can be checked automatically. OpenAI kept o1's internal chain of reasoning largely hidden from users, showing only a summarized version, a decision the company said was meant to preserve the model's ability to reason freely, including exploring incorrect paths, without that unpolished reasoning being taken as a final, endorsed answer. The approach produced substantial accuracy gains on benchmarks requiring multi-step reasoning, including competition mathematics and coding tasks such as SWE-bench.
DeepSeek-R1 and open reasoning models
In January 2025, the Chinese lab DeepSeek released DeepSeek-R1, an open-weights reasoning model trained with large-scale reinforcement learning that matched or approached the performance of OpenAI's reasoning models on several benchmarks at a fraction of the reported training cost, and unlike o1, exposed its full chain of reasoning to users. The release triggered what became known as the DeepSeek market shock, prompting a large sell-off in AI infrastructure stocks and a broader reassessment of assumptions about the compute required to train a frontier-level reasoning model. DeepSeek also released smaller distilled versions of R1, making reasoning-style behavior available in models cheap enough to run on modest hardware.
Test-time compute
The defining tradeoff of reasoning models is that they can convert additional Test-time compute, meaning more computation spent at the moment of answering rather than during training, into higher accuracy on difficult problems. This differs from the traditional scaling relationship described by Scaling laws, which focused on model size, training data, and training compute; reasoning models added a further axis along which capability could be improved without training a new, larger model at all. Researchers including Noam Brown, who worked on reasoning and strategic game-playing systems before joining OpenAI's reasoning model effort, have framed this as analogous to the way a human expert given more time to think produces better answers than one forced to respond instantly.
Evaluation and limits
Reasoning models perform especially well on tasks with objectively checkable answers, which has driven their evaluation toward benchmarks like ARC-AGI and competition mathematics rather than more subjective writing or conversational tasks, where the benefit of extended reasoning is less consistent and the additional inference time and cost are harder to justify. Reasoning models also remain subject to the same Hallucination (AI) risk as other language models, since longer reasoning does not guarantee a factually correct final answer, and can occasionally produce reasoning traces that sound rigorous while embedding an incorrect step.