Reasoning Models: When AI Learned to Think Before Answering

In September 2024, OpenAI released o1 (previewed internally as "Strawberry"), a model that spent compute at inference time rather than producing an immediate answer. Before responding, o1 generated an extended internal chain of reasoning steps, explored alternative approaches, caught and corrected errors in its own logic, and then produced a final answer. The idea of prompting LLMs to "show their work" before answering had been demonstrated in 2022 to substantially improve performance on reasoning tasks. What o1 did differently was make the reasoning internal, variable in length, and learned through reinforcement learning rather than prompted by the user.

The benchmark results were striking. o1 achieved 83% on the 2024 USA Math Olympiad qualifying exam, compared to GPT-4o's 13%. On PhD-level science questions (GPQA Diamond), o1 scored 78% against GPT-4o's 53%. On ARC-AGI, the benchmark designed to resist memorisation, o3 (the December 2024 successor) scored 87.5% in high-compute mode on a task where GPT-4 had achieved near zero. The gap between a model that answered immediately and one permitted to reason for extended periods was measurable, large, and consistent across domains.

The paradigm shift was in where intelligence was located. Pre-o1 AI applied fixed compute per token: more capable models were trained longer on more data, not run longer on each query. o1 demonstrated that variable inference compute, spending more on harder problems at runtime, could substitute for or amplify pre-training scale. A model might not know the answer but could compute its way to it. This created a new design dimension: for simple queries, use a fast cheap model; for hard problems, give a reasoning model more time and budget.

Extended thinking traces, the long chains of reasoning tokens generated before the answer, revealed qualitatively new behaviour. The model appeared to second-guess itself, backtrack from wrong paths, try alternative approaches, and self-correct. This was not simulated deliberation performed for the user's benefit; it was the mechanism through which o1 achieved its scores. It also introduced new failure modes: very long reasoning chains could arrive at wrong answers with high apparent confidence, and the internal reasoning was not always visible or auditable, making failures harder to diagnose.

By early 2025, reasoning models had become a distinct product category. Anthropic's Claude 3.7 Sonnet (February 2025) introduced an extended thinking mode. Google's Gemini 2.0 Flash Thinking launched in January 2025. DeepSeek R1 (January 2025) developed reasoning emergently through reinforcement learning without supervised chain-of-thought data. The question of how to evaluate, trust, and deploy systems whose thinking is extended and partly opaque became the central challenge of the reasoning model era, and remains the frontier of production AI engineering.