Skip to content

Benchmark

MathArena Apex Shortlist

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

Models measured
38
Scores recorded
38
Sources
1
IQ weight
1.63

Score distribution

62.9%
Mean
68.1%
Median
21.7%
Std dev
32.5%
10th percentile
89.4%
90th percentile
8.4%
Min
8.4%98.4%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.63 for chat models, measured over 38 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 5.5xhigh· MathArena data98.4%
2Gemini 3.1 Pro Previewhigh· MathArena data92%
3Kimi K3max· MathArena data92%
4Claude Opus 4.8max· MathArena data90.4%
5DeepSeek V4 Flashmax· MathArena data89.4%
6DeepSeek V4 Promax· MathArena data87.8%
7Claude Opus 4.6high· MathArena data86.7%
8Gemini 3.5 Flashhigh· MathArena data82.5%
9GPT 5.4xhigh· MathArena data81.4%
10GPT 5.2high· MathArena data79.3%
11Kimi K2.6high· MathArena data77.1%
12Step 3.7 Flashhigh· MathArena data76.6%
13GLM 5.1· MathArena data71.8%
14Gemini 3.6 Flash· MathArena data71.3%
15Step 3.5 Flash· MathArena data70.7%
16DeepSeek V3.2 Speciale· MathArena data69.7%
17Gemini 3 Flashhigh· MathArena data68.6%
18GLM 5· MathArena data68.6%
19GLM 5.2max· MathArena data68.1%
20Gemini 3 Pro Previewpreview· MathArena data66.5%
21Claude Opus 4.7xhigh· MathArena data63.8%
22Qwen3.5 397B A17B· MathArena data60.1%
23Grok 4.1 Fastreasoning· MathArena data57.9%
24Grok 4· MathArena data57.5%
25Kimi K2.5high· MathArena data57.5%
26Nemotron 3 Super· MathArena data57.5%
27GPT 5.1high· MathArena data55.9%
28Qwen3.5 27B· MathArena data52.1%
29DeepSeek V3.2think· MathArena data50.5%
30Kimi K2 Thinking· MathArena data47.3%
31GPT OSS 120Bhigh· MathArena data45.2%
32Qwen3.5 35B A3B· MathArena data44.7%
33GPT-5 minihigh· MathArena data39.9%
34Qwen3.5 4B· MathArena data32.5%
35Qwen3.5 9B· MathArena data29.8%
36Qwen3 30B A3Bthink· MathArena data23.4%
37Qwen3 4Bthink· MathArena data16.5%
38Qwen3.5 2B· MathArena data8.4%

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)