Skip to content

Benchmark

MathArena 08/2026

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

Models measured
20
Scores recorded
23
Sources
1
IQ weight
2.04

Score distribution

62.4%
Mean
68.4%
Median
23.5%
Std dev
27.4%
10th percentile
88.1%
90th percentile
11.3%
Min
11.3%93.5%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 2.04 for chat models, measured over 20 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-6.2 Astramax· MathArena data93.5%
2GPT-6.2 Solmax· MathArena data93.5%
3Gemini 4 Carbonhigh· MathArena data88.1%
4GPT-6.1 Solmax· MathArena data86.3%
5GPT-6 Astramax· MathArena data81.9%
6Claude Fable 5.1high· MathArena data79.8%
7GPT-6 Solmax· MathArena data78.6%
8Claude Fable 5.1low· MathArena data78%
9Claude Opus 5.5high· MathArena data77.4%
10GPT-6 Astralow· MathArena data74.7%
11Qwen3.8 Maxmax· MathArena data69.6%
12Muse Wondermax· MathArena data68.4%
13Kimi K3.1 Promax· MathArena data64.3%
14Kimi K3max· MathArena data61.9%
15DeepSeek V4.1 Flash 1017max· MathArena data60.7%
16DeepSeek V4.1 Flashmax· MathArena data54.8%
17Gemini 4 Argonhigh· MathArena data54.2%
18DeepSeek V4.1 Flash 1017max, DSH· MathArena data46.7%
19Grok 4.7xhigh· MathArena data35.1%
20Muse Spark Openxhigh· MathArena data29.2%
21Gemini 3.8 Flashhigh· MathArena data27.4%
22Muse Spark 1.3max· MathArena data20.2%
23Mistral Large 4high· MathArena data11.3%

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)