Teaching models to think

Where the chatbot race was about breadth, the reasoning race was about depth: getting a model to work through a hard problem in steps rather than answer in one pass. It started as a prompting trick — chain-of-thought prompting showed that asking a model to “think step by step” sharply improved its results on maths and logic — and matured into the ReAct pattern of interleaving thought with action.

The paradigm shift was making reasoning a training objective rather than a prompt. A test-time-compute paper argued that letting a model spend more compute thinking could outperform a larger model answering quickly, and OpenAI’s o1 shipped exactly that: a model trained to produce a long internal chain of thought before replying. DeepSeek’s R1 then reproduced the approach openly and cheaply, and Anthropic’s Claude 3.7 and Gemini 2.5 made extended reasoning a standard frontier feature.

The results climbed fast. Reasoning systems from OpenAI and DeepMind reached gold-medal standard at the 2025 International Mathematical Olympiad, and a year later models scored perfect marks. Whether the visible chain of thought is genuine reasoning or an elaborate, legible form of pattern-matching remains contested — but its usefulness, and its cost in inference compute, are not.