OpenAI o1

OpenAI o1 is a reasoning-focused large language model released by OpenAI in September 2024, positioned as the first in a new series trained primarily through reinforcement learning rather than conventional pre-training scaling . The model introduces "reasoning tokens" that enable inference-time scaling—spending more compute during generation to improve output quality on complex problems .

In 云启资本's early assessment, o1 was not a universal replacement for GPT-4o: it excelled at mathematics, coding, and scientific reasoning—reaching "PhD-level" performance in physics and programming per 真格基金's December 2024 transcript —but underperformed on everyday natural language tasks . OpenAI itself labeled the release a "preview," with API head Michelle Pokrass noting it required different prompting patterns and would unlock "entirely new use cases" for developers .

The o1 architecture reportedly combines reinforcement learning with Chain-of-Thought techniques, and industry observers speculated it used process reward models (PRMs) to score intermediate reasoning steps, plus possible Monte Carlo tree search for synthetic training data generation . A significant implication, raised by 峰瑞资本, was that o1 established a new "RL Scaling Law" that could bypass pre-training data limits by shifting GPU allocation from the traditional 9:1:0 ratio (pre-training:post-training:inference) toward 1:1:1 . 真格基金's Yusen Dai later argued this post-training reinforcement learning path, extended through o3, was what "unlocked" viable AI agent capabilities .