Reasoning Models
AI models trained to spend extra computation on a problem before answering — producing internal chains of thought, checking their work, and trading speed for higher accuracy on maths, coding, and logic.
Reasoning models are language models optimised to think before they speak. Instead of predicting the next token immediately, they generate a longer internal deliberation — exploring approaches, backtracking, and verifying results — before producing a final answer.
The pattern became prominent with models such as OpenAI’s o-series and similar systems from other labs. Two ingredients matter:
- Training: reinforcement learning rewards solutions that arrive at correct final answers, teaching the model to produce useful intermediate reasoning.
- Test-time compute: the model is allowed to generate many more tokens at inference time, scaling accuracy with how long it is permitted to think.
The trade-offs are cost and latency. A reasoning model may burn hundreds or thousands of hidden tokens on a question a standard model answers instantly, and it can still confidently get straightforward tasks wrong while overthinking them. Practical systems often route between fast and reasoning models depending on task difficulty, a technique known as model cascading or routing.