Glossary
FrontierMath (tiers 1 to 3)
FrontierMath (tiers 1 to 3), on the Epoch AI leaderboard (CC BY 4.0, latest version we checked), is a set of unpublished research-level math problems, kept private by Epoch. The score is the share of correct answers.
What it measures
FrontierMath (tiers 1 to 3), on the Epoch AI leaderboard (CC BY 4.0, latest version we checked), is a set of unpublished research-level math problems, kept private by Epoch. The score is the share of correct answers.
When it matters
It shows the ceiling of reasoning. Low scores are common on this test. What I watch is the trend across versions of the same family, not a tight finish at the top.
How to read it
The problems are unpublished and kept private by Epoch, on the tiers 1 to 3 leaderboard, in the latest version we checked. Low scores are common. I watch the trend across versions of the same family.
Who leads
- Claude Fable 5.2 Anthropic95.1%
- GPT-6.1 Sol OpenAI93.7%
- GPT-6 Astra OpenAI93.7%
- Claude Opus 5.5 Anthropic91.2%
- GPT-6.1 Terra OpenAI90.9%
- Claude Fable 5.1 Anthropic90.2%
- GPT-6 Sol OpenAI89.8%
- GPT-5.6 Sol OpenAI89.1%
- Claude Sonnet 5.5 Anthropic88.8%
- GPT-5.5 Pro OpenAI87.7%
Source: Epoch AI, latest version we checked.