Skip to content

Benchmark

MathArena Apex

Math benchmarks measure how well models solve maths problems. Scores are percentages. Higher is better.

Models measured
54
Scores recorded
62
Sources
3
IQ weight
2.67

Score distribution

26.7%
Mean
19.8%
Median
26.4%
Std dev
0.52%
10th percentile
65.6%
90th percentile
0%
Min
0%81.3%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 2.67 for chat models, measured over 54 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#Model
1Claude Opus 4.8max· MathArena data81.3%
2GPT 5.5xhigh· MathArena data80.2%
3Hy4 preview· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim74.2%
4Hy4 preview2025 set· Reported by Tencent Hy· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.74.2%
5DeepSeek V4 Flash· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim72.1%
6GPT-5.4 Proxhigh· MathArena data69.8%
7Kimi K3max· MathArena data65.6%
8DeepSeek V4.1 Flash· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim65.6%
9DeepSeek V4.1 FlashDeepSeek V4.1 launch chart · exact public value · setting not further specified· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.65.6%
10DeepSeek V4.1 Flashpass@1 · reasoning_effort 100· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.65.6%
11Kimi K3DeepSeek V4.1 launch chart · exact public value · setting not further specified· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.65.6%
12DeepSeek V4 Pro 0813DeepSeek V4.1 launch chart · exact public value · setting not further specified· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.65.3%
13Gemini 3.1 Pro Previewhigh· MathArena data60.9%
14DeepSeek V4 Flash 0731DeepSeek V4.1 launch chart · exact public value · setting not further specified· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.58.6%
15GPT 5.4xhigh· MathArena data54.2%
16Qwen3.7 Max· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim44.5%
17Claude Opus 4.7xhigh· MathArena data40.6%
18Hy3 (Deep Infra)· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim38.7%
19Hy3 (Deep Infra)Hy3 official Benchmark Appendix · highest reasoning tier unless noted· Reported by Tencent Hunyuan· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.38.7%
20Doubao Seed 2.1 Turbo· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim35.4%
21Claude Opus 4.6high· MathArena data34.5%
22Gemini 3.5 Flashhigh· MathArena data32.3%
23Doubao Seed 2.1 Pro· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim31.3%
24DeepSeek V4 Promax· MathArena data28.1%
25DeepSeek V4 Flashmax· MathArena data27.1%
26Gemini 3.6 Flash· MathArena data26%
27Kimi K2.6high· MathArena data24%
28Gemini 3 Pro Previewpreview· MathArena data23.4%
29Gemini 3 Pro Preview· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim23.4%
30Gemini 3 Pro PreviewGemini 3 Pro Model Card · results as of November 2025· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.23.4%
31GLM 5.2max· MathArena data19.8%
32Gemini 3 Flashhigh· MathArena data15.6%
33Step 3.7 Flashhigh· MathArena data14.6%
34GPT 5.2high· MathArena data13.5%
35Step 3.5 Flash· MathArena data13.5%
36GLM 5.1· MathArena data11.5%
37GLM 5· MathArena data10.9%
38DeepSeek V3.2 Speciale· MathArena data9.4%
39Kimi K2.5high· MathArena data8.9%
40Nemotron 3 Super· MathArena data7.8%
41Grok 4.1 Fastreasoning· MathArena data5.2%
42Qwen3 235B (Web)think· MathArena data5.2%
43Seed 2.0 Mini· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim4.2%
44Qwen3.5 35B A3B· MathArena data4.2%
45DeepSeek V3.2think· MathArena data2.1%
46Grok 4· MathArena data2.1%
47Qwen3 4Bthink· MathArena data2.1%
48Qwen3.5 27B· MathArena data2.1%
49Claude Sonnet 4.5think· MathArena data1.6%
50DeepSeek R1· MathArena data1%

Showing 1–50 of 62 results.

Tokens the model generated per task, as measured by the source. Comparable within this benchmark only.

Data credits: MathArena (ETH SRI / INSAIT), MIT — cite dekoninck2026matharena (licence) · Data by LLM Stats · BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links.