Skip to content

Benchmark

MathArena 01/2026

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

Models measured
24
Scores recorded
27
Sources
1
IQ weight
1.75

Score distribution

57.4%
Mean
57.1%
Median
12.4%
Std dev
44.6%
10th percentile
72.8%
90th percentile
23.9%
Min
23.9%76.1%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.75 for chat models, measured over 24 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 5.4xhigh· MathArena data76.1%
2DeepSeek V4 Promax· MathArena data73.9%
3GPT 5.5xhigh· MathArena data73.9%
4Claude Opus 4.6high· MathArena data72.8%
5Kimi K2.6high· MathArena data71.7%
6Gemini 3.1 Pro Previewhigh· MathArena data70.7%
7GPT 5.2high· MathArena data67.9%
8Gemini 3.1 Pro Previewmedium· MathArena data67.4%
9GLM 5.1· MathArena data65.2%
10Kimi K2.5high· MathArena data62.5%
11Gemini 3 Pro Previewpreview· MathArena data62%
12Step 3.5 Flash· MathArena data60.3%
13Gemini 3 Flashhigh· MathArena data58.2%
14DeepSeek V3.2think· MathArena data57.1%
15Qwen3.5 397B A17B· MathArena data54.9%
16GLM 5· MathArena data54.4%
17Grok 4.1 Fastreasoning· MathArena data53.3%
18Qwen3.5 27B· MathArena data53.3%
19Claude Opus 4.7xhigh· MathArena data52.2%
20Gemini 3.1 Pro Previewlow· MathArena data50%
21Qwen3.5 35B A3B· MathArena data50%
22Nemotron 3 Super· MathArena data48.9%
23GPT 5.2low· MathArena data47.9%
24Qwen3.5 9B· MathArena data44.6%
25Qwen3 30B A3Bthink· MathArena data39.1%
26Qwen3.5 4B· MathArena data38.1%
27Qwen3 4Bthink· MathArena data23.9%

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)