Skip to content

Benchmark

MathArena 04/2026

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

Models measured
23
Scores recorded
23
Sources
1
IQ weight
1.42

Score distribution

34.6%
Mean
25.8%
Median
24.1%
Std dev
8.2%
10th percentile
72.1%
90th percentile
1.9%
Min
1.9%97.5%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.42 for chat models, measured over 23 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-6 Astramax· MathArena data97.5%
2Claude Fable 5.1max· MathArena data77.5%
3GPT 5.5xhigh· MathArena data72.1%
4Kimi K3max· MathArena data59.4%
5Claude Fable 5max· MathArena data54.1%
6Muse Spark 1.3max· MathArena data49.2%
7Qwen3.8 Maxmax· MathArena data45.5%
8DeepSeek V4.1 Flashmax· MathArena data45%
9Claude Opus 4.8max· MathArena data34.4%
10Qwen3.6 35B· MathArena data33.8%
11Gemini 3.8 Flashhigh· MathArena data33.6%
12Gemini 3.1 Pro Previewhigh· MathArena data25.8%
13DeepSeek V4 Flashmax· MathArena data23.8%
14DeepSeek V4 Promax· MathArena data22.1%
15Gemini 3.5 Flashhigh· MathArena data21.3%
16Gemini 3.7 Flashhigh· MathArena data19.3%
17Step 3.7 Flashhigh· MathArena data18.9%
18GLM 5.2max· MathArena data18.4%
19Gemini 3.6 Flash· MathArena data16%
20Step 3.5 Flash· MathArena data13.9%
21GLM 5.1· MathArena data8.2%
22Claude Opus 4.7xhigh· MathArena data4.1%
23Qwen3.5 2B· MathArena data1.9%

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)