Can OpenAI’s New o1 Model Reach Human‑Level Reasoning?

OpenAI has unexpectedly launched the o1 series, a new family of large language models designed for general complex reasoning by generating extensive internal chains of thought before answering.

The o1 models dramatically outperform GPT‑4o on challenging benchmarks: on PhD‑level physics problems GPT‑4o scores 59.5, while o1 reaches 92.8. In the 2024 IOI competition, a fine‑tuned o1 version achieved 213 points (top 49% of human participants) and could surpass the gold‑medal threshold with enough attempts.

Beyond contests, o1 ranks in the top 89% on Codeforces competitive‑programming problems and places among the top 500 U.S. students in the AIME pre‑selection.

Model Variants

Three versions are released:

o1 : the full, most powerful model (currently limited in public release).

o1‑preview : an early version available to ChatGPT Plus and API users.

o1‑mini : a faster, more cost‑effective model for tasks that require reasoning but less world knowledge.

All variants are trained with reinforcement learning to produce a long “thought chain” before output, a process that improves with longer reasoning time and yields a new scaling law for LLM performance.

The models excel in scientific domains: they can annotate cell‑sequencing data, generate complex quantum‑optics formulas, and automate multi‑step workflows for developers.

Human evaluations show o1‑preview outperforms GPT‑4o on reasoning‑intensive categories such as data analysis, coding, and mathematics, though it may not be optimal for some pure language tasks.

OpenAI plans to increase the inference time from seconds to minutes, hours, or even days, aiming for breakthroughs comparable to new drug discovery or solving the Riemann hypothesis.

Access

ChatGPT Plus and Team users will receive early access, with weekly usage limits (30 messages for o1‑preview, 50 for o1‑mini). API access starts with Tier 5 users (those who have spent over $1,000 on the OpenAI API).

The goal is to let future versions think for hours, days, or weeks, accepting higher inference costs for potentially transformative scientific advances.