ML//model//GPT//o1

o1 is OpenAI's first frontier reasoning model (September 2024), the model that proved extended thinking scales. Its core insight is that more test-time compute gives better answers: instead of making the model bigger, it lets the model think longer.

o1 is OpenAI's first frontier reasoning model (September 2024), the model that proved extended thinking scales. Its core insight is that more test-time compute gives better answers: instead of making the model bigger, it lets the model think longer.

OpenAI describes it as trained with reinforcement learning to produce long chains of thought before answering, with accuracy improving both with more training and with more thinking time (test-time compute). The rest is unpublished: process rewards (PRMs, scoring each step) and search over branches (tree search) are plausible readings, never documented facts, and the reasoning it shows varies from run to run in ways that can be measured from outside without revealing the mechanism.

The thinking is hidden from the user: the model generates internal reasoning tokens that are discarded before the response is shown.

It dominated math, coding and science benchmarks, outperforming GPT-4 on GPQA, MATH and competition-level problems.

Its weakness is overthinking: it is empirically worse than GPT-4 on simple common-sense questions, because extended thinking on trivial problems triggers a distributional shift to the wrong basin.

OpenAI did not publish full technical details, but DeepSeek R1 published its pipeline (SFT plus GRPO) and matched o1, which suggests the approach is not magic.