Skip to content

Benchmark

MathArena AIME 2026

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

Models measured
31
Scores recorded
31
Sources
1
IQ weight
2.06

Score distribution

94.7%
Mean
95.8%
Median
3.6%
Std dev
90%
10th percentile
98.3%
90th percentile
82.5%
Min
82.5%100%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 2.06 for chat models, measured over 31 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 Opus 4.8max· MathArena data100%
2GPT 5.5xhigh· MathArena data100%
3GPT 5.4xhigh· MathArena data99.2%
4Gemini 3.1 Pro Previewhigh· MathArena data98.3%
5GPT 5.2high· MathArena data98.3%
6Claude Opus 4.6high· MathArena data96.7%
7DeepSeek V4 Promax· MathArena data96.7%
8Gemini 3 Flashhigh· MathArena data96.7%
9Gemini 3.6 Flash· MathArena data96.7%
10GLM 5· MathArena data96.7%
11Kimi K3max· MathArena data96.7%
12Step 3.5 Flash· MathArena data96.7%
13Claude Opus 4.7xhigh· MathArena data95.8%
14DeepSeek V4 Flashmax· MathArena data95.8%
15GLM 5.1· MathArena data95.8%
16Kimi K2.5high· MathArena data95.8%
17Kimi K2.6high· MathArena data95.8%
18Gemini 3.5 Flashhigh· MathArena data95%
19Step 3.7 Flashhigh· MathArena data95%
20DeepSeek V3.2think· MathArena data94.2%
21Grok 4.1 Fastreasoning· MathArena data94.2%
22Qwen3.5 397B A17B· MathArena data94.2%
23Qwen3.5 35B A3B· MathArena data93.3%
24Qwen3.5 9B· MathArena data92.5%
25Gemini 3 Pro Previewpreview· MathArena data91.7%
26Nemotron 3 Super· MathArena data91.7%
27Qwen3.5 27B· MathArena data91.7%
28GLM 5.2max· MathArena data90%
29Qwen3.5 4B· MathArena data89.7%
30Qwen3 30B A3Bthink· MathArena data88.3%
31Qwen3 4Bthink· MathArena data82.5%

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)