Epoch AI - Weirdml V3: leaderboard

WeirdML v3 as run by Epoch AI: agentic tasks in which a coding agent explores unfamiliar data and builds machine-learning pipelines with unspecified goals and limited feedback. The score combines the area under the score-versus-tokens curve with the final value; higher is better. Not comparable with v2.

Metric: Score. Source: epoch.ai. Saturation forecast: Around December 2027. 26 models tracked. Updated 10 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-6 AstraxHigh42.22#6
2GPT-6.1 SolxHigh38.25#10
3Claude Opus 5.5xHigh31.2#5
4GPT-6.1 TerraxHigh28.09#21
5Claude Fable 5.1xHigh25.97#15
6Gemini 4 ArgonHigh20.26#12
7Claude Sonnet 5.5xHigh19.94#19
8GPT-6 SolxHigh19.73#26
9Claude Opus 5xHigh19.01#20
10Claude Fable 5xHigh15.59#27
11muse-wonder-1.0xHigh15.17
12GPT-5.6 SolxHigh14.99#30
13Claude Haiku 5.5xHigh11.7#70
14Kimi K3.1 ProxHigh9.89#29
15DeepSeek V4.1 Flash (1017)High9.19#41

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=epoch-ai-weirdml-v3 · How It Works · Data refreshed daily, snapshot 2026-11-16.