AI's best score on research-level maths problems went from 6% to 98% in about 14 months
In July 2025 the best AI solved 6% of Epoch AI's hardest, professor-written maths problems. By September 2026 top models scored 97.6–100%.
FrontierMath: hard math problems solved
Best score on Epoch AI's FrontierMath (corrected v2 set), plotted at each model's release date.
What this measures: Hundreds of unpublished problems written by mathematicians. Tiers 1–3 are advanced; Tier 4 is research-level. Answers are checked automatically.
- Tiers 1–3 (advanced)
- Tier 4 (research-level)
Tier 4 (research-level), Saturated: Epoch AI considers Tier 4 saturated: GPT-6 Astra solved the last problem no model had solved (September 2026), and GPT-6.1 Sol scores 100%. Now watch: FrontierMath Erdős (68 genuinely open problems).
Epoch AI re-ran older models on the corrected v2 problem set (June 2026), so early points come from those re-runs. OpenAI funded FrontierMath and has access to many problems; Epoch keeps a held-out set. Answers are checked automatically. These are not proofs of open problems.
Show data table and sources
| Series | Date | Label | Value |
|---|---|---|---|
| Tiers 1–3 (advanced) | 17 Dec 2024 | o1 (high) | 15% |
| Tiers 1–3 (advanced) | 31 Jan 2025 | o3-mini (high) | 19% |
| Tiers 1–3 (advanced) | 16 Apr 2025 | o4-mini (high) | 36% |
| Tiers 1–3 (advanced) | 7 Aug 2025 | GPT-5 (high) | 55% |
| Tiers 1–3 (advanced) | 11 Dec 2025 | GPT-5.2 Pro | 74% |
| Tiers 1–3 (advanced) | 5 Mar 2026 | GPT-5.4 Pro | 83% |
| Tiers 1–3 (advanced) | 23 Apr 2026 | GPT-5.5 Pro | 88% |
| Tiers 1–3 (advanced) | 9 Jul 2026 | GPT-5.6 Sol | 89% |
| Tiers 1–3 (advanced) | 1 Sept 2026 | Claude Fable 5.1 | 90% |
| Tiers 1–3 (advanced) | 3 Sept 2026 | GPT-6 Astra | 94% |
| Tiers 1–3 (advanced) | 9 Nov 2026 | Claude Fable 5.2 (max) | 95% |
| Tier 4 (research-level) | 16 Apr 2025 | o4-mini | 4.9% |
| Tier 4 (research-level) | 7 Aug 2025 | GPT-5 | 22% |
| Tier 4 (research-level) | 11 Dec 2025 | GPT-5.2 Pro | 46% |
| Tier 4 (research-level) | 5 Mar 2026 | GPT-5.4 Pro | 59% |
| Tier 4 (research-level) | 23 Apr 2026 | GPT-5.5 Pro | 78% |
| Tier 4 (research-level) | 9 Jun 2026 | Claude Fable 5 | 90% |
| Tier 4 (research-level) | 3 Sept 2026 | GPT-6 Astra | 98% |
| Tier 4 (research-level) | 29 Sept 2026 | GPT-6.1 Sol (max) | 100% |
| Tier 4 (research-level) | 9 Nov 2026 | Claude Fable 5.2 (max) | 100% |
- Epoch AI benchmark data (CSV export) (Epoch AI)
- FrontierMath Tier 4 benchmark page (Epoch AI)
FrontierMath Tier 4 is Epoch AI’s hardest set of maths problems, written by professors and postdoctoral researchers. Each has one final answer that a computer checks. When Epoch first ran it in July 2025, the best model (o4-mini) solved 6%, 3 of 48 problems.
Scores then climbed steadily: 10% in August 2025, 31% by December 2025 (GPT-5.2 Pro) and 48% in May 2026 (a Google DeepMind “AI co-mathematician” system).
In June 2026 Epoch found errors in 42% of the problems and released a corrected version 2, with 43 problems. On it, Claude Fable 5 scored 90.2% in June 2026. In September, GPT-6 Astra reached 97.6%, solving the last problem no model had solved, and GPT-6.1 Sol scored 100%. Epoch’s brief says this leads it “to consider the benchmark saturated”: the top models now get almost everything right, so the test can no longer show progress.
What it means
Problems that professors find hard are now within reach of top AI when there is a single answer to check. That is not the same as open research, where nobody knows the answer: on Epoch’s set of 68 genuinely open Erdős problems, the best model solves 3. Tracking further progress needs harder tests.
Caveats
- The 6% and 98% come from different versions of the test. Epoch corrected the problems in June 2026, so the two scores are not directly comparable.
- OpenAI funded FrontierMath and has access to many of its problems (28 of the 48 original Tier 4 problems). Epoch keeps a held-out set and reports no significant score gap between held-out and other problems.
- Some answers reached the result through unintended shortcuts (for example, finding a 2011 preprint that skipped most of the intended work).
- Several records were manual web-app runs or best-of-several attempts, not single standard runs.