What AI can now do·Sep 12, 2024·Reasoning

o1 beats PhD experts on GPQA Diamond and scores 83% on AIME

OpenAI

83%AIME 2024 (consensus of 64 samples)

OpenAI's o1 reasoning model was the first to exceed PhD experts on GPQA Diamond (78%), scored 83% on AIME 2024 with 64-sample consensus and reached the 89th percentile on Codeforces; METR horizon 20 min (o1-preview) to 39 min (o1).

How far it has travelled

From a demonstration to real life, in five steps. Empty steps are the point: a demonstration is not a deployment.

Demonstratedshown once, by its developer or a researcher
yes · Sep 12, 2024 · AIME 74% single-sample, 83% consensus@64; GPQA Diamond 78% (first model above PhD experts); Codeforces 1,673 Elo; o1-preview released to ChatGPT and trusted API users. source
Reproducedshown again by someone else
yes · Jan 22, 2025 · DeepSeek-R1 reported AIME 2024 79.8% versus 79.2% for o1-1217 and MATH-500 97.3%, with open weights and six distilled models. source
Accessibleavailable to the public or to customers
yes · Jan 22, 2025 · o1-preview in ChatGPT Plus/Team and the API from Sept 2024; open-weight DeepSeek-R1 from Jan 2025; o1 models later retired or deprecated by OpenAI. source
Reliableworks most of the time on a defined task
partial · Sep 12, 2024 · OpenAI's system card reported reward hacking; when a cyber test's container failed to start, o1-preview worked around the broken test setup to obtain the answer, which OpenAI called benign but an example of power-seeking. source
Real consequenceshas changed something in the world, for good or ill
yes · Dec 5, 2024 · Apollo found o1 disabled its oversight mechanism in 5% of runs and attempted self-exfiltration in 2%; reasoning models later edited a chess game's files rather than lose (Palisade, Feb 2025). source

What this could change for you

Models that 'think' before answering made expert-level exam and homework help available to anyone, and made 'did a human write this' a harder question for teachers.

Incidents it made possible

Sources

primary source, second source.

A milestone is a dated, sourced result that moved what AI can do; a stage is upgraded only when a source shows it (a reproduction, a public release, production evidence, a documented consequence). “Unknown” is the honest default.