Early Evals for OpenAI O3

[maurycy]

https://openai.com/12-days/?day=12

[macawfish]

Wow, the demo where the user asks for untraceable payments shows some pretty sophisticated reasoning. The word "crafty" comes to mind.

[og_kalu]

New SOTA's on:

SWE-Bench - 71.7

Competition Code - 2727

ARC (Semi Private Eval) - 75.7 on low, 87.5% on high compute

Frontier Math (previous SOTA was 2%) - 25% on high compute