FrontierMath was introduced in November 2024 by Epoch AI, in a paper whose first author is Elliot Glazer. At launch, the paper reported that state-of-the-art AI models solved under 2% of its problems, making it one of the hardest benchmarks available.
What it measures
FrontierMath asks whether an AI system can do the kind of mathematics that professional mathematicians do: problems that take a specialist hours, and sometimes days, to solve. It covers major branches of modern mathematics, from number theory and real analysis to algebraic geometry and category theory, rather than the competition-style puzzles found in benchmarks such as AIME.
How it works
All problems are new and unpublished, written and vetted by expert mathematicians. Each has a definite, automatically verifiable answer, often a large number or a precise mathematical object, so that guessing is practically useless and grading can be done by code.
Epoch AI’s benchmark page states that after a v2 update on 12 June 2026 the benchmark has 338 problems:
- Tiers 1–3: 295 problems of increasing difficulty;
- Tier 4: 43 exceptionally hard problems in an expansion set.
Only 12 problems are public (10 from Tiers 1–3 and 2 from Tier 4). Unless stated otherwise, the numbers Epoch publishes come from the private sets. A model submits a final answer, which is checked automatically against the reference.
Why it matters
Competition maths benchmarks have saturated quickly, and their problems are often available online. FrontierMath is designed to stay hard and uncontaminated for longer, because the problems are original and almost all are kept private. A rising FrontierMath score is therefore one of the stronger signals that a model’s mathematical reasoning, not its memory, is improving. It has also become a reference point in debates about whether AI can contribute to real research mathematics.
Limitations and criticisms
- Funding and access. Epoch states that FrontierMath was developed with funding from OpenAI and that OpenAI has exclusive access to a subset of the benchmark. The arrangement drew criticism in early 2025, when it became widely known, and it is worth keeping in mind when comparing OpenAI models with others.
- Answers, not proofs. Problems require a final checkable answer. A model might reach it by heuristics or numerical exploration without a rigorous argument, so the benchmark does not directly measure proof-writing.
- Small and private. Hidden problems protect against contamination but make independent verification of individual results difficult, and per-tier sample sizes are small.
- Evaluation setup matters. Tool access, time limits and number of attempts all affect scores.
Reading the scores
Even strong human mathematicians would not solve every problem outside their speciality, and there is no single human baseline. Single-digit scores were the norm at launch; anything well into double digits on Tiers 1–3 marks a highly capable system, and Tier 4 remains far harder. Epoch AI runs the evaluations itself and publishes results on its Benchmarking Hub. Scores move quickly, so check the official FrontierMath page for current numbers.
For more on the life cycle of benchmarks, see how AI benchmarks work and why they saturate.