Tests AI models on fresh, hard math problems taken from new competitions and research papers, so the answers cannot have been memorised.
Leaderboard
Top 3 right now
1stGemini 4 Carbon (high)94.28% ± 3.64%
2ndGPT-6.1 Sol (max)92.81% ± 3.10%
3rdGemini 4 Argon (high)84.39% ± 4.65%
What it measures
Models are tested on new math competitions (such as AIME, HMMT and olympiads) shortly after they take place, scored by final-answer accuracy or human grading of proofs.
How to read the numbers
Accuracy on ArXivMath (questions from recent research papers), % with 95% CI — higher is better.
Results
8 rows · higher is better (Accuracy (± 95% CI))CSV
1
Gemini 4 Carbon (high)
Google
94.28% ± 3.64%
$5.08
2
GPT-6.1 Sol (max)
OpenAI
92.81% ± 3.10%
$0.96
3
Gemini 4 Argon (high)
Google
84.39% ± 4.65%
$3.74
4
Claude-Opus-5.5 (high)
Anthropic
82.61% ± 4.72%
$1.93
5
Kimi K3.1 Pro (max)
Moonshot AI
58.27% ± 6.84%
$2.94
6
DeepSeek-V4.1-Flash-1017 (max, DSH)
DeepSeek
42.48% ± 6.32%
$0.33
7
Inkling 2 Small (0.99)
Thinking Machines Lab
31.88% ± 5.14%
$0.60
8
Muse Spark Open (xhigh)
Meta AI
27.02% ± 5.61%
$1.03
Note: Table is the ArXivMath overall board (most recently updated). MathArena has many other boards (final-answer comps, proof comps, Project Euler, etc.).
Source: MathArena (SRI Lab ETH Zurich / INSAIT) · matharena.ai · results as of 2026-11-13 · retrieved 2026-11-17