MMODELYST
Evals/AIME
EVL

AIME

Olympiad math

Source ↗
benchmarkMathAccuracy (% of problems solved exactly)30 items
What this measures

The American Invitational Mathematics Examination — elite high-school competition problems with integer answers from 0–999. A favorite for measuring real reasoning models.

Sample
Find the number of ordered pairs of positive integers (a, b) such that a·b + 63 = 20·lcm(a,b) + 12·gcd(a,b).
Leaderboard · 99 models
Accuracy (% of problems solved exactly) — higher is better
1GPT-5.2
99
2GPT-5 Codex
98.7
3Gemini 3 Flash Preview
97
4DeepSeek V3.2 Speciale
96.7
5MiMo-V2-Flash
96.3
6Gemini 3 Pro Preview
95.7
7GPT-5.1 Codex
95.7
8GLM-4.7
95
9Kimi K2 Thinking
94.7
10KAT-Coder-Pro V1
94.7
11GPT-5
94.3
12GPT-5.1
94
13gpt-oss-120b
93.4
14Grok 4
92.7
15DeepSeek V3.2
92
16GPT-5.1 Codex mini
91.7
17Claude Opus 4.5
91.3
18Qwen3 235B A22B 2507
91
19NVIDIA Nemotron 3 Nano 30B A3B
91
20o4-mini
90.7
21K-EXAONE
90.3
22Grok 4 Fast
89.7
23DeepSeek V3.1 Terminus
89.7
24Nova 2.0 Omni
89.7
25Ring-1T
89.3
26Grok 4.1 Fast
89.3
27gpt-oss-20b
89.3
28Nova 2.0 Pro Preview
89
29Nova 2.0 Lite
88.7
30Qwen3 VL 235B A22B
88.3
31o3
88.3
32Claude 4.5 Sonnet
88
33Apriel-v1.6-15B-Thinker
88
34INTELLECT-3
88
35DeepSeek V3.2 Exp
87.7
36Gemini 2.5 Pro
87.7
37Apriel-v1.5-15B-Thinker
87.5
38o3-mini
87.3
39GLM-4.6
86
40GLM-4.6V
85.3
41GPT-5 mini
85
42ERNIE 5.0 Thinking Preview
85
43Seed-OSS-36B-Instruct
84.7
44Qwen3 VL 32B
84.7
45Grok 3 mini Reasoning
84.7
46Qwen3 Next 80B A3B
84.3
47GPT-5 nano
83.7
48Ring-flash-2.0
83.7
49Claude 4.5 Haiku
83.7
50Qwen3 4B 2507
82.7
51MiniMax-M2.1
82.7
52Qwen3 Max Thinking (Preview)
82.3
53Qwen3 VL 30B A3B
82.3
54Qwen3 235B A22B
82
55Magistral Medium 1.2
82
56Qwen3 Max
80.7
57GLM-4.5-Air
80.7
58Claude 4.1 Opus
80.3
59Motif-2-12.7B-Reasoning
80.3
60Magistral Small 1.2
80.3
61EXAONE 4.0 32B
80
62Falcon-H1R-7B
80
63Doubao Seed Code
79.3
64Mi:dm K 2.5 Pro Preview
78.7
65MiniMax-M2
78.3
66Gemini 2.5 Flash Preview (Sep '25)
78.3
67K2-V2
78.3
68Olmo 3.1 32B Think
77.3
69Mi:dm K 2.5 Pro
76.7
70Llama Nemotron Super 49B v1.5
76.7
71DeepSeek R1 0528 (May '25)
76
72Qwen3 Max (Preview)
75
73NVIDIA Nemotron Nano 12B v2 VL
75
74Claude 4 Sonnet
74.3
75Qwen3 Omni 30B A3B
74
76GLM-4.5
73.7
77Olmo 3 32B Think
73.7
78Claude 4 Opus
73.3
79Gemini 2.5 Flash
73.3
80Qwen3 32B
73
81GLM-4.5V
73
82Cogito v2.1
72.7
83Qwen3 VL 30B A3B Instruct
72.3
84Qwen3 30B A3B
72.3
85Qwen3 235B A22B 2507 Instruct
71.7
86Ling-1T
71.3
87Qwen3 VL 235B A22B Instruct
70.7
88Olmo 3 7B Think
70.7
89Hermes 4 - Llama-3.1 405B
69.7
90NVIDIA-Nemotron-Nano-9B-v2
69.7
91Gemini 2.5 Flash-Lite Preview (Sep '25)
68.7
92Hermes 4 - Llama-3.1 70B
68.7
93Qwen3 VL 32B Instruct
68.3
94DeepSeek R1 (Jan '25)
68
95Qwen3 30B A3B 2507 Instruct
66.3
96Qwen3 Next 80B A3B Instruct
66.3
97Ling-flash-2.0
65.3
98DeepSeek R1 0528 Qwen3 8B
63.7
99Llama 3.1 Nemotron Ultra 253B v1
63.7
Benchmark health
saturating

Top models cluster near the ceiling — differences here are getting too small to mean much. Weight newer, harder benchmarks more.

96.4
top-10 mean
99
best
99
models
Recent changes
GPT-5 mini90.7 → 8522d ago
Nova 2.0 Lite94.3 → 88.729d ago
What improves this score

Datasets & environments shown to raise it.

—+30.0
NuminaMath CoT+6.5
Research using this benchmark
all →
Weak-to-Strong Generalization via Direct On-Policy DistillationVibeThinker-3B: Exploring the Frontier of Verifiable Reasoning in Small Language ModelsKVarN: Variance-Normalized KV-Cache Quantization Mitigates Error Accumulation in Reasoning TasksStale but Stable: Staleness-Adaptive Trust Regions for Stabilizing Asynchronous Reinforcement LearningShare More, Search Less: Collaborative Parallel Thinking for Efficient Test-Time Scaling

Matched by name in title/abstract.