MATH (Level 5) leaderboard
MATH Level 5 leaderboard: accuracy on the hardest tier of competition math problems for each LLM, plotted against API price. Current leader: GPT-5 at 98.13 (% accuracy).
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
OPENOPEN
OPENOPEN
OPENOPEN
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 MATH (Level 5) measures
MATH Level 5 is the hardest of five human-labeled difficulty tiers in the MATH dataset, introduced by Dan Hendrycks and coauthors (UC Berkeley) in "Measuring Mathematical Problem Solving With the MATH Dataset" (NeurIPS 2021). The full dataset holds 12,500 competition-style problems (7,500 for training, 5,000 held out for testing) drawn from contests such as AMC 10/12 and AIME across seven subject areas from prealgebra to precalculus; Level 5 isolates only the hardest slice of that pool, typically requiring multi-step algebraic manipulation or proof-like reasoning rather than a single formula lookup. Models respond in free-response form, ending with a line of the form "ANSWER: $ANSWER" — a boxed answer command is explicitly not required — and the headline score comes from an LLM-judged equivalence check (rather than a brittle literal string match) on the extracted answer, though stricter string-match and symbolic-equivalence scorers are also computed separately.
Epoch AI, whose independent re-run of this benchmark most leaderboards (including this one) draw on, evaluates each model across 8 repeated runs using zero-shot chain-of-thought prompting, each model's API-default temperature, and an empty system prompt, restricted to chat/instruct variants rather than base models; the published score is the average accuracy across those runs. MATH Level 5 is also one of more than 50 evaluations feeding Epoch's broader Capabilities Index, where its raw score is rescaled before being blended with other benchmarks.
How to read this page
This page's cost-vs-quality chart uses each model's output price per 1M tokens as the cost axis, standing in for a per-task dollar figure that MATH Level 5 does not publish. That proxy fits this benchmark reasonably well because Level 5 problems are output-heavy: the input is a single, short competition-math problem statement, but a passing attempt requires a long worked derivation and, for reasoning models, an even longer hidden chain-of-thought that providers bill as output tokens before the final answer appears. Output pricing therefore tracks what it actually costs to run this benchmark far more closely than input pricing would.
Limitations
MATH Level 5 is now widely regarded as close to saturated for frontier reasoning models: Epoch AI noted in April 2025 that o3 and o4-mini both reached 98% accuracy, leaving little headroom to distinguish top systems and making small score gaps hard to interpret as real capability differences. Contamination is a standing concern because the MATH test set has been publicly downloadable since 2021 and OpenAI has confirmed mixing MATH-derived training data into GPT-4's pretraining, so exact or near-duplicate Level 5 problems may have entered later models' pretraining or fine-tuning corpora. Results are also scaffold-sensitive — small changes in prompts or evaluation settings can shift scores meaningfully — and Epoch does not offer public transcript/log access for this specific benchmark, citing ongoing copyright discussions around the source competition problems, which limits independent auditing of exactly what was asked and how answers were scored.