Reasoning Models: OpenAI o1 and After

In September 2024, OpenAI previewed o1, a model that spent extra compute before answering, generating long internal chains of reasoning. On hard math, coding, and science benchmarks, o1 leapt past GPT-4o, sometimes by double-digit points on tests where progress had been crawling.

The core idea was not prompt-level chain of thought. It was training the model, often through reinforcement learning, to use its own thinking tokens effectively, and then letting it spend as many of those tokens as needed at inference time.

What reasoning models do well

• Multi-step math and proofs where intermediate errors compound
• Competitive programming problems requiring search
• Scientific reasoning on benchmarks like GPQA Diamond
• Agentic tasks that benefit from planning and reflection

Competitors followed quickly. Google's Gemini 2.0 Flash Thinking, DeepSeek's R1 in early 2025, and Anthropic's extended thinking mode all adopted variants of the paradigm. Some published training recipes openly; others kept them secret.

“We've developed a new series of AI models designed to spend more time thinking before they respond.”

The big idea: reasoning models reopened the scaling frontier by moving compute from training time into inference time. A model that can think longer is a different kind of model.