OpenAI o1

OpenAI o1 is a language model series launched in September 2024, built on a new training paradigm that departs from the GPT line. Its core innovation is **test-time compute scaling**: the model spends seconds to minutes internally reasoning through problems via a hidden chain-of-thought process, testing approaches and self-correcting before answering — what OpenAI co-founder Greg Brockman described as shifting from "System I" to "System II" thinking .

The series debuted with **o1-preview** and **o1-mini**, products of the internal "Strawberry" project . On benchmarks, o1 reached **83%** on the International Mathematical Olympiad qualifying exam against GPT-4o's 13%, and placed in the top 500 of competitive Codeforces programming . It also outperformed GPT-4o on 54 of 57 MMLU subcategories and scored 78.2% on multimodal MMMU tests .

Industry analysis, including a 30,000-word DeepMind researcher interview published by ZhenFund, frames o1 as combining reinforcement learning with chain-of-thought to unlock doctoral-level reasoning in physics, math, and coding . The model may still use linear autoregressive decoding rather than tree search, based on API token-timing experiments , and its training approach has been likened to AlphaGo Zero's self-play methods .