Skip to content

Benchmark

MathArena 02/2026

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

Models measured
26
Scores recorded
28
Sources
1
IQ weight
0.45

Score distribution

23.8%
Mean
19.4%
Median
17.4%
Std dev
6.5%
10th percentile
37.9%
90th percentile
2.8%
Min
2.8%75.8%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.45 for chat models, measured over 26 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-5.4 Proxhigh· MathArena data75.8%
2GPT 5.5xhigh· MathArena data68.2%
3Gemini 3.1 Pro Previewlow· MathArena data40.6%
4GPT 5.4xhigh· MathArena data37.9%
5GPT 5.2high· MathArena data37.5%
6Claude Opus 4.8max· MathArena data35.5%
7Grok 4.1 Fastreasoning· MathArena data32%
8Qwen3.5 27B· MathArena data31.3%
9Nemotron 3 Super· MathArena data30.5%
10Qwen3.5 35B A3B· MathArena data30.5%
11Qwen3.5 9B· MathArena data27.3%
12GPT 5.2xhigh· MathArena data25.8%
13Qwen3 30B A3Bthink· MathArena data22.7%
14Qwen3.5 4B· MathArena data19.4%
15Gemini 3.1 Pro Previewhigh· MathArena data18.6%
16Qwen3 4Bthink· MathArena data18%
17DeepSeek V4 Flashmax· MathArena data17.7%
18DeepSeek V4 Promax· MathArena data13.3%
19Step 3.7 Flashhigh· MathArena data12.1%
20GLM 5· MathArena data11.7%
21Kimi K2.6high· MathArena data11.7%
22Step 3.5 Flash· MathArena data11.3%
23GLM 5.2max· MathArena data10.5%
24GLM 5.1· MathArena data9.3%
25Gemini 3.5 Flashhigh· MathArena data6.5%
26Claude Opus 4.7xhigh· MathArena data4%
27Claude Opus 4.6high· MathArena data3.2%
28Kimi K2.5high· MathArena data2.8%

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)