FrontierMath Tier 4 leaderboard

What FrontierMath Tier 4 measures

Tier 4 is the hardest slice of Epoch AI's FrontierMath benchmark: 43 exceptionally difficult, research-level problems added as an expansion to the 295-problem base set (Tiers 1-3). Like the rest of FrontierMath, the problems are original and unpublished, written and vetted by expert mathematicians, and each has a single exact answer that the model returns from a Python function so it can be checked automatically. Only two Tier 4 problems are public.

Tier 4 exists because the base tiers started to fill up. A leaderboard that reports a model at 90% cannot show which of two strong models is better, so Epoch kept a separate, harder set. The score here is the percentage of the 43 problems solved at each model's best reasoning effort.

How to read this page

Read Tier 4 as a stress test of the strongest reasoning settings, not as a measure of everyday ability. The cost chart uses output price per 1M tokens, since the work is long chains of reasoning and the prompt is short. The most useful comparison is usually not the top spot but the gap between the leaders and cheaper or open-weight models, which is far wider on Tier 4 than on easier benchmarks. If your workload is not research-grade mathematics, a score here says little about whether a cheaper model will do the job.

Limitations

With 43 problems, one extra solve moves a score by more than two percentage points, so small gaps are meaningless. The best systems on this leaderboard are at or near the ceiling, which limits how much more the set can separate them. As with the base set, the data was revised on 2026-06-12 to fix errors, FrontierMath received funding from OpenAI (which has exclusive access to a subset), and exact-answer grading cannot tell a sound derivation from a lucky guess.