AIME (OTIS Mock) leaderboard
AIME (OTIS Mock) leaderboard: percent-correct on contamination-resistant competition math problems for each LLM, plotted against API price. Current leader: GPT-6 Sol at 100 (% correct).
Cost vs quality
Each dot is a model. The dashed line is the value frontier — top-left is the best trade-off (high score, low cost).
Full leaderboard
OPEN
OPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPEN
OPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPEN
OPENOPENOPEN
OPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENOPENScores from Epoch AI, updated —. Cost is joined from our daily API pricing where a model matches; models we don't price still appear with their score. Always verify on the benchmark's own site before relying on a number.
What AIME (OTIS Mock) measures
AIME (OTIS Mock) evaluates a model's ability to solve competition-level mathematics problems from the OTIS Mock AIME — an annual, unofficial 3-hour, 15-problem exam written by students in Evan Chen's Olympiad Training for Individual Study (OTIS) program to emulate the real American Invitational Mathematics Examination (AIME), typically 2–4 problems harder. Epoch AI's dataset pools 45 problems from three sittings (Mock AIME 2024, 2025 I, and 2025 II); as on the real AIME, every answer is an integer between 0 and 999, enabling exact-match grading instead of subjective judgment. Models are prompted to show their work and finish with "ANSWER: X"; since September 2025 Epoch AI grades with a model-based answer extractor that pulls out only the stated final number — without ever seeing the correct answer — before checking it against the answer key. The three problems that include a diagram are given to models as the underlying Asymptote source code rather than an image, so text-only models aren't disadvantaged.
The reported metric is % correct across the pooled 45-problem set (with a confidence interval, since 45 items is a small sample), and Epoch AI positions the benchmark deliberately between MATH Level 5 (easier) and FrontierMath (harder) on its difficulty ladder, using it as one of the evaluations feeding its broader Epoch Capabilities Index.
How to read this page
Because OTIS Mock AIME problems are proof-style competition math, a model — especially a reasoning model — typically burns hundreds to thousands of output tokens working through algebra, casework, or number-theory steps before it commits to a single final integer, while the problem statement itself is only a short paragraph of input; Epoch AI has even shown that giving a model a larger "thinking budget" changes its score here. Because the task is so output-heavy, this page's cost-vs-quality chart uses each model's output price per 1M tokens as the cost proxy — for a long-reasoning math task like this one, output tokens are where the real per-query spend accumulates.
Limitations
The pool is small — only 45 problems across three exam sittings — so scores carry wide error bars, meaning a few flipped answers can reorder the leaderboard. The benchmark is also showing real ceiling effects: Epoch AI has said the set is "nearly saturated," with a single unsolved problem remaining across all evaluated models as of late 2025, which caps its usefulness for separating current frontier systems. Contamination risk is lower than the real AIME by design — Epoch adopted these problems specifically because the official AIME's decades of archived exams and solutions are widely suspected to have leaked into training data — but the OTIS mock exams themselves are posted publicly with full solutions every year, so the same risk grows with each newer model's training cutoff. Results are also scaffold- and infrastructure-sensitive: Epoch's own write-up documents API timeouts and an output-token cap during one model's evaluation that caused failed requests to be auto-scored as incorrect — a sign that reported numbers reflect evaluation-harness choices as well as raw capability.