{"chart":{"title":"Share of FrontierMath problems solved correctly by AI models","subtitle":"The FrontierMath benchmark tests AI models on research-level mathematics problems that take expert mathematicians hours or days to solve.","note":"This chart shows performance on Tiers 1-3 of the FrontierMath benchmark (295 problems).","citation":"Epoch AI (2026)","originalChartUrl":"https://ourworldindata.org/grapher/ai-frontiermath-over-time?csvType=full&useColumnShortNames=false","selection":["o1 (Dec 2024), high","o4-mini (Apr 2025), high","GPT-5 (Aug 2025), high","GPT-5.2 (Dec 2025), xhigh","GPT-6 Astra, max","Claude Opus 4.8, max","Claude Fable 5.1, max","Gemini 3.1 Pro","Claude Opus 4.1 (Aug 2025), 32K","Kimi K3, max","GPT-4o (Aug 2024)","Qwen3.7-Max"]},"columns":{"Share of FrontierMath problems solved correctly by AI models":{"titleShort":"Share of problems solved correctly by each model","titleLong":"Share of problems solved correctly by each model","descriptionShort":"FrontierMath evaluates models on 295 difficult, research-level problems in advanced mathematics (Tiers 1–3), which can take expert mathematicians hours or days to work through.","descriptionKey":"- This indicator shows the share of FrontierMath problems that AI models solve correctly, based on Epoch AI's evaluation.\n- FrontierMath is a set of 338 original math problems written by experts, covering many areas of advanced mathematics. Many problems are difficult enough that human specialists might need hours or days to solve them.\n- The benchmark has four difficulty tiers. This indicator shows accuracy on Tiers 1–3 (295 problems). Tier 4 contains 43 exceptionally difficult problems and is not included here.\n- In June 2026, Epoch AI reissued FrontierMath after addressing errors in 42% of its problems. This indicator uses the corrected problem set. Scores on it are not comparable with scores on the original set, where the flawed problems held every model well below 100%; models evaluated on both scored around 12 percentage points higher on the corrected set.\n- Scoring is all-or-nothing: models get 1 point for a correct final answer and 0 for anything else, with no partial credit. Models submit their answers as Python code and can use Python while working on problems. This means scores reflect math ability with access to computational tools, not just pen-and-paper reasoning.\n- Only 12 of the problems are publicly available: 10 from Tiers 1–3 and 2 from Tier 4. They are published so researchers can inspect how evaluations work, not to report scores.\n- FrontierMath was developed by Epoch AI with funding from OpenAI, whose GPT models are among those evaluated on this benchmark. OpenAI has exclusive access to a subset of the problems.","shortUnit":"%","unit":"%","timespan":"","type":"Numeric","owidVariableId":1306024,"shortName":"mean_score","lastUpdated":"2026-09-07","nextUpdate":"2026-11-27","citationShort":"Epoch AI (2026) – with minor processing by Our World in Data","citationLong":"Epoch AI (2026) – with minor processing by Our World in Data. “Share of problems solved correctly by each model” [dataset]. Epoch AI, “Epoch AI Benchmark Data” [original data].","fullMetadata":"https://api.ourworldindata.org/v1/indicators/1306024.metadata.json"}},"dateDownloaded":"2026-10-27","activeFilters":{}}