Benchmarks › MATH Level 5

MATH Level 5

Solve the hardest level of competition mathematics questions
98.1top score · GPT-5 (2025-08-07)
35models ranked
23open-weights models
Accuracy (model_graded_equiv on default plots)higher is better
Oct 20, 2026last updated

How the test works

1

Task

1,324 level-5 test questions

  • What the model doesSolve the hardest level of competition mathematics questions
  • Kinds of tasksCompetition mathematics
2

Run

Step-by-step prompt requiring final ANSWER: $ANSWER line

3

Score

Accuracy (model_graded_equiv on default plots)

  • Graded byGemini-1.5-flash-002 checks answer equivalence; other scorer variants remain distinct
  • Better isHigher
Sources: epoch.ai

MATH Level 5 leaderboardtop 10 of 35

Full leaderboard with filters →
ModelMATH Level 5Per taskPriceSpeedSupportsContext
98.1high
—$1.25 / $10—400K
97.8high
—$0.25 / $2—400K
97.8high
—$2 / $8—200K
R1 0528DeepSeekOpen
96.6unstated
—$0.39 / $1.63—128K
95.2medium
—$0.05 / $0.40—400K
89.9
—$0.70 / $0.80—131K
87.3
—$0.40 / $1.60—1M
86.9unstated
—$1 / $5—200K
83
—$2 / $8—1M
DeepSeek V3 0324DeepSeekOpen
75.5unstated
—$0.24 / $0.90—128K

MATH Level 5 questions

What is MATH Level 5?›

Solve the hardest level of competition mathematics questions. It has 1,324 level-5 test questions. Created by MATH dataset authors; Level 5 evaluation by Epoch AI.

Which model is best on MATH Level 5?›

GPT-5 (2025-08-07) leads MATH Level 5 with 98.1, ahead of GPT-5 Mini (2025-08-07) (97.8) and o3 (2025-04-16) (97.8).

What is the best open-weights model on MATH Level 5?›

R1 0528 with 96.6, then DeepSeek R1 Distill Llama 70B (89.9).

How is MATH Level 5 scored?›

Accuracy (model_graded_equiv on default plots); gemini-1.5-flash-002 checks answer equivalence; other scorer variants remain distinct.