Skip to content

Benchmark

MathArena AIME 2025

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

Models measured
41
Scores recorded
49
Sources
1
IQ weight
2.82

Score distribution

77.1%
Mean
87.5%
Median
23.8%
Std dev
47.2%
10th percentile
95.8%
90th percentile
3.3%
Min
3.3%100%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 2.82 for chat models, measured over 41 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#Model
1GPT 5.2high· MathArena data100%
2GPT 5.2xhigh· MathArena data100%
3Step 3.5 Flash· MathArena data98.3%
4Gemini 3 Flashhigh· MathArena data97.5%
5GLM 5· MathArena data96.7%
6DeepSeek V3.2 Speciale· MathArena data95.8%
7Kimi K2.5high· MathArena data95.8%
8Gemini 3 Pro Previewpreview· MathArena data95%
9GPT 5high· MathArena data95%
10DeepSeek V3.2think· MathArena data94.2%
11GPT 5.1high· MathArena data94.2%
12GLM 4.5· MathArena data93.3%
13Grok 4· MathArena data92.5%
14Kimi K2 Thinking· MathArena data92.5%
15DeepSeek V3.2 Expthink· MathArena data91.7%
16GLM 4.6· MathArena data91.7%
17o4 Minihigh· MathArena data91.7%
18DeepSeek V3.1think· MathArena data90.8%
19GPT OSS 120Bhigh· MathArena data90%
20DeepSeek R1· MathArena data89.2%
21GPT OSS 20bhigh· MathArena data89.2%
22Grok 4.1 Fastreasoning· MathArena data89.2%
23OpenAI o3high· MathArena data89.2%
24Gemini 2.5 Pro· MathArena data88.3%
25GPT-5 minihigh· MathArena data87.5%
26o3 Minihigh· MathArena data86.7%
27GPT-5 nanohigh· MathArena data85%
28Claude Sonnet 4.5think· MathArena data84.2%
29o4 Minimedium· MathArena data84.2%
30GLM 4.5 Air· MathArena data83.3%
31Grok 3 Minihigh· MathArena data81.7%
32o1medium· MathArena data81.7%
33Qwen 3 235B A22B· MathArena data80.8%
34o3 Minimedium· MathArena data76.7%
35Gemini 2.5 Flashthinking· MathArena data70.8%
36Claude Opus 4think· MathArena data70%
37Qwen3 30B A3B· MathArena data70%
38QwQ 32B· MathArena data65.8%
39Grok 3 Minilow· MathArena data65%
40o4 Minilow· MathArena data61.7%
41Gemini 2.0 Flashthinking· MathArena data53.3%
42Claude Sonnet 3.7think· MathArena data49.2%
43o3 Minilow· MathArena data48.3%
44GPT-5 nanolow· MathArena data47.2%
45QwQ 32B Preview· MathArena data33.3%
46Gemini 2.0 Flash· MathArena data27.5%
47DeepSeek V3· MathArena data25%
48GPT 4o· MathArena data11.7%
49Claude 3.5 Sonnet· MathArena data3.3%

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)