FrontierMath Benchmark: Scores, Tier 4 and Limits
Quick answer: FrontierMath is a mathematics benchmark created by Epoch AI in November 2024 to test advanced AI reasoning. It uses original, expert-written problems, from difficult undergraduate math to research-level challenges, each with a single answer a computer can check. When checked on November 22, 2026, the top listed scores were 95% on Tiers 1-3 and 100% on Tier 4 (version 2, private sets). Separate projects, FrontierMath: Open Problems and FrontierMath: Erdős, track AI progress on questions that were unsolved when chosen. A perfect score on one set does not mean AI can solve every math problem.
Results checked November 22, 2026 (Epoch’s leaderboards were last updated November 19).
When AI systems began acing school tests and olympiad problems, researchers needed a harder yardstick. FrontierMath was built to be it: hundreds of new, unpublished problems, each written and checked by professional mathematicians, many needing hours or days of expert work. This guide explains what the benchmark tests, what its scores actually measure, and why the frontier has moved on to unsolved problems.
Latest FrontierMath Results
| Collection | Size and scope | Result (checked Nov 22, 2026) | What the number means |
|---|---|---|---|
| Tiers 1-3 (v2) | 295 problems; 10 public, 285 private | Top listed score 95% | Accuracy of the best system on the private set |
| Tier 4 (v2) | 43 problems; 2 public, 41 private | Top listed score 100% | Accuracy on the 41 private problems |
| Open Problems | 49 currently listed | 5 solved by AI, 5 by humans with AI, 39 unsolved | A running catalog, not one model’s score |
| Erdős (launched Sept 1, 2026) | 68 conjectures from 65 Erdős problems | 2 of 68 at launch; 17 solved since, by any means | Answers must be proved in Lean |
A key detail for reading the table: Epoch reports scores on the private sets, so “100% on Tier 4” means 41 of 41 private problems, not the 2 public examples.
What Is FrontierMath?
FrontierMath was created by Epoch AI, a research group that tracks AI progress, and released on November 8, 2024 with a paper by Elliot Glazer, Ege Erdil, Tamay Besiroglu and colleagues. Its design has three features:
- New problems. They were written for the benchmark to reduce the risk that models saw the exact questions during training.
- Automatic checking. Each problem has a definite answer, such as a large integer or an exact expression, and answer formats make random guessing very unlikely to succeed.
- Mostly unpublished. Only 12 of the 338 version 2 problems are public.
What Is FrontierMath Tier 4?
| Tier | Level |
|---|---|
| 1 | Difficult undergraduate and olympiad-style problems, with computation allowed |
| 2 | Advanced graduate mathematics |
| 3 | Problems like early PhD research |
| 4 | Exceptionally difficult research-level problems written by professors and postdocs |
Tier 4 was added as an expansion in 2025, after the original launch, to stay ahead of rapidly improving models. Top systems reached 100% on its version 2 private set by November 2026.
How Are FrontierMath Scores Checked?
For Tiers 1-4, a model submits a final answer, often by running code, and a script compares it with the stored answer. That checks the answer, not every step of the reasoning. FrontierMath: Erdős is stricter: a model must give a proof in the Lean proof language (see how Lean checks proofs), and Open Problems uses verifier programs that test whether a proposed construction really works.
Reported scores can differ between sources because of the benchmark version, public versus private problems, tools allowed, reasoning effort and budget. Always check which setting a number comes from.
What Changed in Version 2?
On June 12, 2026, Epoch released version 2. It corrected 123 problems in Tiers 1-3 and 12 in Tier 4, and removed 5 and 7 problems respectively, so 147 of the previous 350 problems (42%) were affected. The collection now has 338 problems. Scores before and after the update are not directly comparable.
How Fast AI Scores Rose
At launch, the best models solved under 2% of problems in Epoch’s first evaluation. In about two years, the top listed scores reached 95% on Tiers 1-3 and 100% on Tier 4. For the strongest systems, accuracy on these sets is approaching its ceiling under the reported conditions, which is why attention has shifted to unsolved problems. A similar jump happened at the math olympiad; see AI at the Math Olympiad.
Open Problems and FrontierMath: Erdős
Once AI could answer hard questions with known answers, the sharper test became questions nobody has answered. FrontierMath: Open Problems collects research questions that were open when added and can be checked by a program. As of November 22, 2026, its catalog of 49 lists 5 as solved by AI and 5 as solved by humans working with AI. Being unsolved tests originality; it does not automatically mean a problem is harder than a Tier 4 question.
FrontierMath: Erdős, announced September 1, 2026, has 68 conjectures selected by Thomas Bloom from the Erdős problems list. At launch, the best model resolved 2 of 68 (about 3%). Since then, 17 of the 68 have been solved, many outside Epoch’s official runs, such as in OpenAI’s October 2026 release of research proofs.
Who Funded FrontierMath, and Who Can Access It?
Epoch states that “FrontierMath was developed with funding from OpenAI, who has exclusive access to a subset of the benchmark.” This was not disclosed when the benchmark launched; it became public around OpenAI’s o3 announcement in December 2024, and Epoch published a fuller clarification in January 2025, apologizing for the lack of transparency. Epoch keeps private holdout problems for independent evaluation. Access alone does not show that a model trained on the problems. FrontierMath: Open Problems is developed independently, supported by Schmidt Sciences.
How FrontierMath Compares With Other Tests
| Evaluation | Main task | How it is checked |
|---|---|---|
| MATH (2021) | Competition problems | Final answer |
| FrontierMath Tiers 1-4 | Hard problems with known answers | Final answer, by script |
| FrontierMath: Open Problems | Previously unsolved questions | Verifier program |
| FrontierMath: Erdős | Prove or disprove a conjecture | Lean proof |
| First Proof | Write research-level proofs | Expert review |
Percentages from these tests should not be combined into one “best AI at math” chart; they measure different things.
What a Score Does and Doesn’t Mean
- It shows success on the tested problems under the stated settings.
- It does not show that a model can choose good questions, build theories or explain why a result is true.
- It does not guarantee that your next question will be answered correctly; see Can AI Do Math?
Watch: FrontierMath Explained by Its Creators
If you learn better by watching, these videos are worth your time.
Check Your Understanding
Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.
Answer each question and press Submit to see your result right away.
Frequently Asked Questions
What is FrontierMath?
FrontierMath is a benchmark created by Epoch AI in November 2024 to test advanced AI math reasoning. It uses original, expert-written problems with answers a computer can check, from difficult undergraduate level to research level. Most problems are not public.
What is the highest score on FrontierMath?
When checked on November 22, 2026, Epoch AI listed top scores of 95% on FrontierMath Tiers 1-3 and 100% on Tier 4 (version 2, private sets). Scores change often, so check the date and setting.
What is FrontierMath Tier 4?
The hardest tier: exceptionally difficult research-level problems written by professors and postdocs. Version 2 has 43 problems, 41 of them private. Top systems reached 100% on the private set by November 2026.
Did OpenAI fund FrontierMath?
Yes. Epoch says OpenAI funded the original benchmark and has exclusive access to a subset of it. The funding was not disclosed at launch and became public around December 2024; Epoch clarified the arrangement in January 2025.
What are FrontierMath Open Problems?
A catalog of research questions that were unsolved when added, checked by verifier programs. As of November 22, 2026, 5 of the 49 listed were solved by AI and 5 by humans working with AI.
Keep Exploring
- Can AI Do Math? What to Trust in 2026
- AI at the IMO: Gold in 2025, Reported 42/42 in 2026
- Erdős Problems Solved by AI: What’s Really New?
Sources and Further Reading
- Epoch AI: FrontierMath hub
- Epoch AI: FrontierMath Tiers 1-3 (v2) leaderboard
- Epoch AI: FrontierMath Tier 4 (v2) leaderboard and changelog
- Epoch AI: FrontierMath Open Problems
- Epoch AI: FrontierMath Erdős
- Glazer et al., FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI (arXiv 2411.04872)
- Wikipedia: FrontierMath
Related to This Article
More math articles
- How to Solve a Quadratic Equation by Completing the Square?
- The Best Grade 5 ELA Practice Tests for Missouri Students
- Rounding Numbers Up to the Millions
- Top 10 SSAT Middle-Level Math Practice Questions
- How to Help Your 4th Grade Student Prepare for the Florida FAST Math Test
- How is the AFOQT Test Scored?
- The CIO, Labor’s Breakthrough and the New Deal’s Exclusions
- Conic Sections
- The Best Grade 7 ELA Practice Tests for North Dakota Students
- Stability and change in opinion


What people say about "FrontierMath Benchmark: Scores, Tier 4 and Limits - Effortless Math"?
No one replied yet.