MATH Leaderboard

High-school competition math problems. · Prices checked November 17, 2026

The MATH benchmark uses high-school competition problems that need multi-step symbolic working rather than recall. It separates models that can hold a chain of reasoning together from those that pattern-match toward an answer.

Leaderboard

60 of 103 scored

LeaderboardScores as published by each benchmark. Prices checked November 17, 2026.

Gemini 2.0 Flash leads on MATH at 90.9%. The median across the 103 models we hold a score for is 68%. The cheapest model on this board is Qwen2.5-3B-Instruct at $0.0500/M blended, scoring 65.9%.

#ModelProviderScoreInput /1MOutput /1MBlended 70/30Context
1Gemini 2.0 FlashGoogle90.9%$0.1000$0.4000$0.19001,000,000
2Gemini 1.5 ProGoogle86%$1.25$5.00$2.382,000,000
3Qwen2.5-MaxQwen85%$1.60$6.40$3.0432,000
4Qwen2.5-72B (Groq)Groq83.1%$0.7900$0.7900$0.7900128,000
5Qwen2.5-72B (Together)Together AI83.1%$1.20$1.20$1.2032,000
6Qwen2.5-72B (fw)Fireworks AI83.1%$0.9000$0.9000$0.900032,000
7Qwen2.5-72B (DI)DeepInfra83.1%$0.3500$0.4000$0.3650128,000
8Qwen2.5-72B (Hyp)Hyperbolic83.1%$0.4000$0.4000$0.4000128,000
9Qwen2.5-72B-InstructQwen83.1%$0.3600$0.4000$0.3720128,000
10Qwen2.5-32B-InstructQwen83.1%$0.7000$0.7000$0.7000128,000
11Command ACohere80%$2.50$10.00$4.75256,000
12Phi-4Microsoft80%$0.0700$0.1400$0.091016,000
13Qwen2.5-14B-InstructQwen80%$0.3500$0.3500$0.3500128,000
14Claude Sonnet 3.5Anthropic78.3%$3.00$15.00$6.60200,000
15Gemini 1.5 FlashGoogle78%$0.0750$0.3000$0.14251,000,000
16MiniMax-Text-01MiniMax77.4%$0.7000$2.80$1.331,000,000
17Llama 3.3 70BMeta77%$0.2300$0.4000$0.2810128,000
18Llama 3.3 70B (Groq)Groq77%$0.5900$0.7900$0.6500128,000
19Llama 3.3 70B (Together)Together AI77%$0.8800$0.8800$0.8800128,000
20Llama 3.3 70B (DI)DeepInfra77%$0.2300$0.4000$0.2810128,000
21Llama-3.3-70B (Cerebras)Cerebras77%$0.5900$0.9900$0.7100128,000
22Yi-Lightning01.AI76.4%$0.1400$0.1400$0.140016,000
23Grok 2xAI76.1%$2.00$10.00$4.40131,000
24Grok 2 VisionxAI76.1%$2.00$10.00$4.4032,000
25Nova ProAmazon76%$0.8000$3.20$1.52300,000
26GPT-4oOpenAI75.9%$2.50$10.00$4.75128,000
27Qwen2.5-7B-InstructQwen75.5%$0.1000$0.1000$0.1000128,000
28GLM-4-PlusZhipu AI74.2%$7.00$7.00$7.00128,000
29GLM-4VZhipu AI74.2%$7.00$7.00$7.002,000
30Llama 3.1 405BMeta73.8%$0.8000$0.8000$0.8000128,000
31Llama 3.1 405B (Groq)Groq73.8%$2.99$2.99$2.99128,000
32Llama 3.1 405B (Together)Together AI73.8%$3.50$3.50$3.50128,000
33Llama 3.1 405B (fw)Fireworks AI73.8%$3.00$3.00$3.00131,000
34Llama 3.1 405B (Rep)Replicate73.8%$9.50$9.50$9.50128,000
35Llama 3.1 405B (DI)DeepInfra73.8%$0.8000$0.8000$0.8000128,000
36Llama-3.1-405B (Cerebras)Cerebras73.8%$0.9900$0.9900$0.9900128,000
37Llama 3.1 405B (Hyp)Hyperbolic73.8%$4.00$4.00$4.00128,000
38GPT-4 TurboOpenAI73.4%$10.00$30.00$16.00128,000
39Hunyuan-TurboTencent73%$3.50$7.00$4.55256,000
40GPT-4o miniOpenAI70.2%$0.1500$0.6000$0.2850128,000
41Gemini 2.0 Flash LiteGoogle70%$0.0750$0.3000$0.14251,000,000
42Mistral Large 2Mistral70%$2.00$6.00$3.20128,000
43Mistral-Large-2 (NIM)Nvidia70%$2.00$6.00$3.20128,000
44Doubao-Pro-256kByteDance70%$0.5600$1.40$0.8120256,000
45Doubao-Pro-128kByteDance70%$0.5600$1.40$0.8120128,000
46Claude Haiku 3.5Anthropic69.2%$0.8000$4.00$1.76200,000
47Llama 3.1 70BMeta68%$0.3500$0.4000$0.3650128,000
48Llama 3.2 90B VisionMeta68%$0.8800$0.8800$0.8800128,000
49Llama 3.1 70B (Groq)Groq68%$0.5900$0.7900$0.6500128,000
50Llama 3.1 70B (DI)DeepInfra68%$0.3500$0.4000$0.3650128,000
51Llama 3.1 70B (Hyp)Hyperbolic68%$0.4000$0.4000$0.4000128,000
52Hermes-3-70B (Hyp)Hyperbolic68%$0.4000$0.4000$0.4000128,000
53Llama 3.1 70B (Anyscale)Anyscale68%$1.00$1.00$1.00128,000
54Llama 3.1 70B (Lepton)Lepton AI68%$0.8000$0.8000$0.8000128,000
55Llama 3.1 70B (DB)Databricks68%$1.00$3.00$1.60128,000
56Gemma 3 27BGoogle67%$0.0800$0.3500$0.1610128,000
57Qwen2.5-3B-InstructQwen65.9%$0.0500$0.0500$0.050032,000
58Phi-4 MiniMicrosoft64%$0.0400$0.1600$0.0760128,000
59Llama 4 MaverickMeta61.2%$0.3500$1.00$0.5450128,000
60Yi-Large01.AI60%$3.00$3.00$3.0032,000

48 of the 60 scores shown link to the page they were read on (underlined; hover for publisher, variant and date read). The rest have no recorded source yet and are being worked through.

About these scores

  • High-school competition math problems.
  • Scores are as published by the model’s provider or a public leaderboard. Provider-reported figures are self-reported and are not independently re-run by us.
  • Where a hosted variant of an open-weights model has no score of its own, it inherits the base model’s, because the weights are the same.
  • Prices beside each score are the provider’s published token rates. How the blended rate is calculated.

Showing the top 60 of 103 models we hold a MATH score for.

Frequently asked questions

What is the MATH benchmark?

A set of competition mathematics problems spanning algebra, geometry, number theory and precalculus. Each requires a worked solution rather than a lookup, and grading checks the final answer, so partial reasoning that lands on the wrong result scores nothing.

Other leaderboards

Rankings that weigh price too

Leaderboards are recomputed on every deploy from the catalogue. Prices checked November 17, 2026.