MMODELYST
Evals/SciCode
EVL

SciCode

benchmarkCodeaccuracy %
Leaderboard · 100 models
accuracy % — higher is better
1Claude Fable 5
60.2
2Gemini 3.1 Pro Preview
58.9
3Muse Spark 1.1
58.2
4GPT-5.4
56.6
5Gemini 3 Pro Preview
56.1
6GPT-5.6 Sol (max)
56.1
7GPT-5.5
56.1
8GPT-5.6 Sol
56
9GPT-5.2 Codex
54.6
10Claude Opus 4.7
54.5
11Grok 4.5
54.1
12GPT-5.6 Terra (max)
53.9
13Claude Sonnet 5
53.6
14Kimi K2.6
53.5
15Claude Opus 4.8
53.5
16GPT-5.3 Codex
53.2
17Gemini 3.5 Flash
53.1
18GPT-5.6 Luna (max)
52.5
19GPT-5.2
52.1
20Claude Opus 4.6
51.9
21GPT-5.6 Terra
51.6
22Muse Spark
51.5
23Gemini 3 Flash Preview
50.6
24GLM-5.2 (max)
50.5
25GPT-5.5 Instant (May 2026)
50.3
26MiMo-V2.5-Pro
50.2
27Grok Build 0.1 0616
50.2
28DeepSeek V4 Pro
50
29GPT-5.6 Luna
50
30GPT-5.4 mini
49.9
31Claude Opus 4.5
49.5
32Kimi K2.5
49
33Qwen3.7 Max
48.8
34GPT-5.5 Instant (May 2026)
48.6
35Kimi K2.7 Code
47.5
36Grok 4.3
47.3
37MiniMax-M2.7
47
38Qwen3.6 Max Preview
46.9
39GPT-5.4 nano
46.9
40Claude Sonnet 4.6
46.8
41o4-mini
46.5
42GLM-5
46.2
43Grok 4
45.7
44Grok 4.20 0309 v2
45.6
45Qwen3.7 Plus
45.5
46MiniMax-M3
45.4
47GLM-4.7
45.1
48DeepSeek V4 Flash
44.9
49Grok 4.20 0309
44.7
50Claude 4.5 Sonnet
44.7
51Grok 4.1 Fast
44.2
52Grok 4 Fast
44.2
53DeepSeek V3.2 Speciale
44
54GLM-5.1
43.8
55GLM-5-Turbo
43.6
56GLM 5V Turbo
43.5
57Gemma 4 31B
43.4
58GPT-5.1
43.3
59Claude 4.5 Haiku
43.3
60Qwen3 Max Thinking
43.1
61MiMo-V2.5
43.1
62GPT-5
42.9
63Gemini 2.5 Pro
42.8
64Nova 2.0 Pro Preview
42.7
65MiniMax-M2.5
42.6
66GPT-5.1 Codex mini
42.6
67MiMo-V2-Pro
42.5
68Qwen3 235B A22B 2507
42.4
69Kimi K2 Thinking
42.4
70Ring-2.6-1T
42.4
71Qwen3.5 122B A10B
42
72Qwen3.5 397B A17B
42
73Gemini 3.1 Flash-Lite
41.9
74Nex-N2-Pro
41.8
75Gemini 2.5 Pro Preview (May' 25)
41.6
76Hy3-preview
41.2
77GPT-5 mini
41
78Cogito v2.1
41
79o3
41
80Claude 4.1 Opus
40.9
81GPT-5 Codex
40.9
82MiniMax-M2.1
40.7
83Doubao Seed Code
40.7
84Qwen3.6 Plus
40.7
85Grok 3 mini Reasoning
40.6
86DeepSeek V3.1 Terminus
40.6
87Gemini 2.5 Flash Preview (Sep '25)
40.5
88Qwen3.5 Omni Plus
40.5
89Step 3.5 Flash
40.4
90GPT-4.1 mini
40.4
91DeepSeek R1 0528 (May '25)
40.3
92Claude 3.7 Sonnet
40.3
93GPT-5.1 Codex
40.2
94Claude 4 Sonnet
40
95Step 3.7 Flash
40
96Gemma 4 26B A4B
40
97Qwen3 VL 235B A22B
39.9
98Qwen3 235B A22B
39.9
99o3-mini
39.9
100Nemotron 3 Ultra 550B A55B
39.9
Benchmark health
headroom

Plenty of room at the top — this benchmark still separates frontier models clearly.

56.7
top-10 mean
60.2
best
100
models
Recent changes
GLM-5.2+36.16d ago
JT-4.1 Flash 236B A21B+38.26d ago
Grok 4.5+54.16d ago
Muse Spark 1.1+58.26d ago
GPT-5.6 Luna (max)+52.56d ago
What improves this score

Datasets & environments shown to raise it.

No data yet.
Research using this benchmark
all →
Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B AgentBenchEvolver: Frontier Task Synthesis via Solution-Centric EvolutionResearchLoop: An Evidence-Gated Control Plane for AI-Assisted Research

Matched by name in title/abstract.