AI · Math Benchmarks

AI Math Reasoning (Mock AIME)

Competition mathematics: from under 1% to 100% in three years.

Saturated — A hard problem became a solved problem: the frontier reached 100% in April 2026.

• GPT-4’s 0.6% → GPT-5.5 Pro’s 100%
• o1-mini was the reasoning inflection
• Mock set: real AIME leaks into training
• Frontier = best mean score to date

Epoch AI runs the frontier models against the OTIS Mock AIME set: competition problems written to match the American Invitational Mathematics Examination without reusing its questions. The mock set matters, because real AIME problems leak into training data and make published AIME scores unreliable. The frontier went from GPT-4’s 0.6% in March 2023 to GPT-5.5 Pro’s 100% in April 2026. The largest step came in September 2024, when o1-mini took the record from 9.7% to 46.9%, trained to reason step by step before answering. Reaching 100% from there took another nineteen months. Each point is the highest mean score any model had reached at that date, taken from Epoch’s published run data. Epoch changed how it grades this set in September 2025, from a model shown the answer key to one that only extracts the final answer for an exact match, and nine of the twenty points, o1-mini’s among them, carry only the earlier grades. At 100% the set is saturated and cannot record what a stronger model would do.

Source: Epoch AI benchmark runs (OTIS Mock AIME 2024-2025)

Download the data: ai_math.csv · ai_math.json

All curves · The chart is drawn in your browser and needs JavaScript. Every number it plots is in the table above.