>>105957270
It looks like they don't include the reasoning traces but you can see from the outputs themselves that the model is way less verbose than typical LLMs.
For example:
https://github.com/aw31/openai-imo-2025-proofs/blob/main/problem_3.txt
>So universal constant c can be taken 4. So c<=4.
>Need sharpness: show can't be less.

It's plausible that the reasoning is also more conservative with word usage. It turns out the secret to gold-level math skills was training it to talk like Kevin.