Skip to content

Benchmark

MathArena 09/2026

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

Models measured
16
Scores recorded
17
Sources
1
IQ weight
1.46

Score distribution

41.6%
Mean
39.6%
Median
22.6%
Std dev
17%
10th percentile
71.7%
90th percentile
1.9%
Min
1.9%80.5%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.46 for chat models, measured over 16 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#Model
1Claude Fable 5.2max· MathArena data80.5%
2GPT-6.2 Astramax· MathArena data79.3%
3GPT-6.2 Solmax· MathArena data71.7%
4Gemini 4 Carbonhigh· MathArena data62.3%
5DeepSeek V4.1 Flash 1017max· MathArena data56%
6DeepSeek V4.1 Flash 1017max, DSH· MathArena data49.7%
7Claude Opus 5.5high· MathArena data47.2%
8Muse Wondermax· MathArena data41.5%
9Gemini 4 Argonhigh· MathArena data39.6%
10GPT-6 Astramax· MathArena data37.7%
11GPT-6.1 Solmax· MathArena data35.9%
12Grok 4.7xhigh· MathArena data27.7%
13Muse Spark Openxhigh· MathArena data27.7%
14Gemini 3.8 Flashhigh· MathArena data26.7%
15Kimi K3.1 Promax· MathArena data17%
16DeepSeek V4.1 Flashmax· MathArena data5.7%
17Qwen3.8 Maxmax· 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)