Skip to content

Benchmark

MathArena HMMT Feb 2026

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

Models measured
31
Scores recorded
31
Sources
1
IQ weight
2.01

Score distribution

87.9%
Mean
89.4%
Median
9.3%
Std dev
78.8%
10th percentile
97%
90th percentile
53%
Min
53%98.5%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 2.01 for chat models, measured over 31 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 data98.5%
2GPT 5.4xhigh· MathArena data97.7%
3GPT 5.2high· MathArena data97%
4Kimi K3max· MathArena data97%
5Claude Opus 4.6high· MathArena data96.2%
6Claude Opus 4.8max· MathArena data95.5%
7Gemini 3.5 Flashhigh· MathArena data95.5%
8Gemini 3.1 Pro Previewhigh· MathArena data94.7%
9Kimi K2.6high· MathArena data94.7%
10Claude Opus 4.7xhigh· MathArena data93.9%
11DeepSeek V4 Flashmax· MathArena data93.9%
12DeepSeek V4 Promax· MathArena data93.9%
13GLM 5.2max· MathArena data92.4%
14Gemini 3 Flashhigh· MathArena data89.4%
15Gemini 3.6 Flash· MathArena data89.4%
16GLM 5.1· MathArena data89.4%
17Qwen3.5 397B A17B· MathArena data87.9%
18Step 3.7 Flashhigh· MathArena data87.9%
19Kimi K2.5high· MathArena data87.1%
20Gemini 3 Pro Previewpreview· MathArena data86.4%
21GLM 5· MathArena data86.4%
22Grok 4.1 Fastreasoning· MathArena data86.4%
23Step 3.5 Flash· MathArena data86.4%
24Nemotron 3 Super· MathArena data84.8%
25DeepSeek V3.2think· MathArena data84.1%
26Qwen3.5 35B A3B· MathArena data81.8%
27Qwen3.5 27B· MathArena data81.1%
28Qwen3 30B A3Bthink· MathArena data78.8%
29Qwen3.5 4B· MathArena data72.9%
30Qwen3.5 9B· MathArena data71.2%
31Qwen3 4Bthink· MathArena data53%

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)