Benchmark

WeirdML (v2)

Measures: Unusual machine-learning engineering tasks (shape recognition, digit classification, chess outcomes and others) where the model writes PyTorch code that is trained and scored automatically.

Maintainer
Håvard Tveit Ihle
Introduced
2025
Status
Active
Saturation
Saturated
Typical provenance
Independent
In scoring set
No

Official leaderboard

Leaderboard

Independent and third-party results only.

Leaderboard
ModelAverage accuracyProvenanceSourceEvaluated
GPT-6 Astra
OpenAI · Promax effort
93.6%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-6 Astra
OpenAI · Max effort
93.3%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Fable 5.1
Anthropic · Max effort
92.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-6 Astra
OpenAI · High effort
92.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Fable 5.1
Anthropic · High effort
92.3%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Fable 5
Anthropic · Max effort
91.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Opus 5
Anthropic · Max effort
91.8%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Opus 5
Anthropic · High effort
91.6%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5.6 Sol
OpenAI · Promax effort
89.4%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5.6 Sol
OpenAI · High effort
88.8%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Fable 5
Anthropic · High effort
87.8%Third partyWeirdML Leaderboard via Epoch AI—
GPT-5.6 Sol
OpenAI · Max effort
87.0%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Opus 586.3%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5.5
OpenAI · Xhigh effort
84.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5.5
OpenAI · High effort
83.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Opus 4.8
Anthropic · Xhigh effort
82.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Kimi K3
Moonshot AI · Max effort
82.6%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5.3-Codex79.3%Third partyWeirdML Leaderboard via Epoch AI—
GPT-5.6 Terra
OpenAI · High effort
78.3%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Opus 4.6
Anthropic · High effort
78.0%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Opus 4.6
Anthropic · Unknown effort
77.9%Third partyWeirdML Leaderboard via Epoch AI—
GPT-5.3-Codex
OpenAI · Xhigh effort
77.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5.4
OpenAI · Xhigh effort
77.7%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Opus 4.7
Anthropic · High effort
76.4%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Opus 4.776.4%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Opus 4.7
Anthropic · Unknown effort
76.4%Third partyWeirdML Leaderboard via Epoch AI—
Claude Opus 4.8
Anthropic · Medium effort
76.0%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Opus 4.7
Anthropic · Max effort
75.5%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GLM-5.3
Z.ai (Zhipu AI) · Max effort
75.4%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5.2
OpenAI · Xhigh effort
72.2%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Gemini 3.1 Pro72.1%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Opus 4.8
Anthropic · None effort
70.5%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GLM-5.2
Z.ai (Zhipu AI) · Max effort
70.1%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Gemini 3 Pro69.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Sonnet 5
Anthropic · High effort
68.8%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GLM-5.2
Z.ai (Zhipu AI) · High effort
67.3%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Grok 4.6
SpaceXAI (xAI) · High effort
67.3%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5.5
OpenAI · None effort
67.2%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
DeepSeek-V4-Pro
DeepSeek · Max effort
66.2%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Sonnet 4.6
Anthropic · Medium effort
66.1%Third partyWeirdML Leaderboard via Epoch AI—
Claude Opus 4.665.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Opus 4.5
Anthropic · 16K effort
63.7%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5.2
OpenAI · Medium effort
63.4%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
DeepSeek-V4-Flash
DeepSeek · Max effort
63.0%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Gemini 3.5 Flash
Google DeepMind · High effort
62.6%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Gemini 3 Flash61.6%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5.6 Luna
OpenAI · High effort
60.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5.1
OpenAI · High effort
60.8%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5
OpenAI · High effort
60.7%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5 Pro
OpenAI · High effort
60.4%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5.4 mini
OpenAI · High effort
60.3%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Muse Spark 1.2
Meta · Xhigh effort
60.3%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
o3-pro
OpenAI · High effort
58.2%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5.4
OpenAI · None effort
57.4%Third partyWeirdML Leaderboard via Epoch AI—
GPT-5.4 Pro
OpenAI · None effort
57.4%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GLM-5.157.1%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
DeepSeek-V4-Flash
DeepSeek · High effort
57.0%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Gemini 3.6 Flash
Google DeepMind · High effort
56.1%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Kimi K2.655.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5-Codex54.5%Third partyWeirdML Leaderboard via Epoch AI—
GPT-5-Codex
OpenAI · High effort
54.5%Third partyWeirdML Leaderboard via Epoch AI—
Kimi K2.7 Code54.1%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Gemini 2.5 Pro
Google DeepMind · 16K effort
54.0%Third partyWeirdML Leaderboard via Epoch AI—
GPT-5 mini
OpenAI · High effort
52.7%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
o4-mini
OpenAI · High effort
52.6%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
o3
OpenAI · High effort
52.4%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Gemma 4 31B52.3%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Grok 4.2052.3%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Gemini 3.1 Flash-Lite52.2%Third partyWeirdML Leaderboard via Epoch AI—
Grok 4.3
SpaceXAI (xAI) · Unknown effort
49.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5.2
OpenAI · None effort
49.6%Third partyWeirdML Leaderboard via Epoch AI—
GPT-5.2
OpenAI · Low effort
49.6%Third partyWeirdML Leaderboard via Epoch AI—
GPT-5.4 nano
OpenAI · High effort
49.2%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
DeepSeek-V4-Pro
DeepSeek · Max effort
48.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GLM-548.2%Third partyWeirdML Leaderboard via Epoch AI—
gpt-oss-120b
OpenAI · High effort
48.2%Third partyWeirdML Leaderboard via Epoch AI—
gpt-oss-120b
OpenAI · High effort
48.2%Third partyWeirdML Leaderboard via Epoch AI—
Claude Sonnet 4.5
Anthropic · 16K effort
47.7%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
o1-preview47.6%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
DeepSeek-V3.2-Speciale46.7%Third partyWeirdML Leaderboard via Epoch AI—
Claude Sonnet 4.546.7%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
DeepSeek-V4-Pro
DeepSeek · High effort
46.5%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Grok 4.5
SpaceXAI (xAI) · Unknown effort
46.4%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Sonnet 4
Anthropic · 16K effort
46.1%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
o1
OpenAI · High effort
46.1%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Opus 4.1
Anthropic · 16K effort
45.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Grok 445.7%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
DeepSeek-V4-Flash
DeepSeek · Max effort
45.6%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Kimi K2.545.6%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Haiku 4.545.4%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Haiku 4.5
Anthropic · 16K effort
44.1%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Sonnet 443.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
DeepSeek-V4-Flash
DeepSeek · High effort
43.8%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude Opus 4
Anthropic · 16K effort
43.7%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Mistral Medium 3.543.7%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
o3-mini
OpenAI · High effort
43.7%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Mercury 243.2%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Grok 4 Fast42.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Kimi K2 Thinking42.8%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Grok 3 Mini
SpaceXAI (xAI) · High effort
42.6%Third partyWeirdML Leaderboard via Epoch AI—
Grok 3 Mini
SpaceXAI (xAI) · High effort
42.6%Third partyWeirdML Leaderboard via Epoch AI—
Gemini 2.5 Flash
Google DeepMind · 16k effort
41.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
gpt-oss-120b
OpenAI · Medium effort
41.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
DeepSeek-R141.6%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Qwen3-Coder-480B-A35B-Instruct41.2%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Qwen3-235B-A22B-Thinking-250741.0%Third partyWeirdML Leaderboard via Epoch AI—
Gemini 2.5 Flash
Google DeepMind · 16K effort
41.0%Third partyWeirdML Leaderboard via Epoch AI—
gpt-oss-20b
OpenAI · High effort
40.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GLM-4.5
Z.ai (Zhipu AI) · Thinking effort
40.6%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude 3.5 Sonnet40.0%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5 Instant (gpt-5-chat-latest)39.8%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Qwen3.5-27B39.5%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
DeepSeek-V3.2-Exp
DeepSeek · Thinking effort
39.5%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-4.539.4%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Kimi K239.4%Third partyWeirdML Leaderboard via Epoch AI—
GPT-4.139.0%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Gemini 3.5 Flash-Lite
Google DeepMind · High effort
39.0%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Qwen3-235B-A22B-Instruct-250738.7%Third partyWeirdML Leaderboard via Epoch AI—
Qwen3-235B-A22B-Thinking-250738.7%Third partyWeirdML Leaderboard via Epoch AI—
DeepSeek-V3.1
DeepSeek · Thinking effort
38.4%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5 nano
OpenAI · High effort
38.1%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5.4 nano
OpenAI · None effort
38.0%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-5.4 mini
OpenAI · None effort
37.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-4.1 mini37.6%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
DeepSeek-V3.137.5%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Qwen3-235B-A22B37.3%Third partyWeirdML Leaderboard via Epoch AI—
Qwen3-235B-A22B-Thinking-250737.3%Third partyWeirdML Leaderboard via Epoch AI—
Grok 337.2%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
MiniMax-M2.737.0%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
gpt-oss-20b
OpenAI · Medium effort
36.8%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Kimi K236.7%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
DeepSeek-R136.5%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
o1-mini
OpenAI · Medium effort
36.3%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Gemini 2.5 Flash-Lite
Google DeepMind · 16K effort
35.2%Third partyWeirdML Leaderboard via Epoch AI—
Gemma 4 26B A4B35.2%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Grok Code Fast 135.1%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Qwen3.6-35B-A3B34.5%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Qwen3-Coder-Next34.4%Third partyWeirdML Leaderboard via Epoch AI—
Mistral Medium 3.133.1%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Inkling
Thinking Machines Lab · High effort
32.3%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude 3.5 Sonnet31.0%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude 3.5 Haiku30.7%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Qwen3-30B-A3B29.8%Third partyWeirdML Leaderboard via Epoch AI—
GPT-5 nano
OpenAI · Low effort
25.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Gemini 2.0 Flash25.8%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-4o25.1%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Gemini 1.5 Flash24.9%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Llama 4 Maverick24.5%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Grok 222.2%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Gemini 1.5 Pro22.2%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Llama 3.1 405B21.4%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude 3 Opus19.2%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-4.1 nano19.0%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-4 Turbo18.0%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Qwen2.5-72B-Instruct16.0%Third partyWeirdML Leaderboard via Epoch AI—
Llama 3.3 70B14.4%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-412.4%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-4o mini11.8%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Qwen2-72B-Instruct11.3%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude 3 Sonnet10.2%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude 3 Haiku9.8%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Llama 3.1 70B9.0%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Claude 2.17.1%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
GPT-3.5 Turbo3.5%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Mixtral 8x22B3.2%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
Llama 3.1 8B1.7%Third partyhttps://htihle.github.io/weirdml.html via Epoch AI—
  • Epoch AI, 'Capabilities & benchmarking', epoch.ai/benchmarks, CC BY 4.0.

Sources