EnigmaForge: leaderboard
Procedurally generated, mechanically verified reasoning benchmark in which the question is hidden inside a story: the model must reconstruct a hidden world from narrative clues and reach the right decision. Scored by task success over 600 instances, with missing or invalid answers counted as failures; higher is better (EnigmaForge, 2026).
Metric: Task success (%) over 600 procedurally generated story instances, all items (missing or invalid answers count as failures). Source: robottwo.github.io. Saturation forecast: Around February 2027. 33 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6.1 Sol (Low) | 84.5 | #10 (GPT-6.1 Sol) |
| 2 | Gemini 4 Argon (Low) | 81.2 | #11 (Gemini 4 Argon) |
| 3 | GPT-6 Astra (Low) | 80.3 | #6 (GPT-6 Astra) |
| 4 | GPT-6.1 Terra (Low) | 77.2 | #20 (GPT-6.1 Terra) |
| 5 | GPT-6 Sol (Low) | 74.2 | #25 (GPT-6 Sol) |
| 6 | Gemini 3.7 Flash (Low) | 68.7 | #32 (Gemini 3.7 Flash) |
| 7 | GPT-5.6 Sol (Low) | 68.5 | #30 (GPT-5.6 Sol) |
| 8 | Claude Fable 5.2 (Low) | 68 | #1 (Claude Fable 5.2) |
| 9 | Gemini 3.8 Flash (Low) | 59 | #36 (Gemini 3.8 Flash) |
| 10 | GLM-5.3 (Low) | 53.2 | #66 (GLM-5.3) |
| 11 | Gemini 4 Carbon (Low) | 50.8 | #2 (Gemini 4 Carbon) |
| 12 | GPT-5.6 Terra (Low) | 50.7 | #62 (GPT-5.6 Terra) |
| 13 | DeepSeek V4.1 Flash (Low) | 48.7 | #59 (DeepSeek V4.1 Flash) |
| 14 | Muse Wonder (Low) | 45.5 | #28 (Muse Wonder) |
| 15 | DeepSeek V4.1 Flash (1017) (Low) | 44.2 | #39 (DeepSeek V4.1 Flash (1017)) |
Interactive version: theaggregate.ai/benchmark?slug=enigmaforge · How It Works · Data refreshed daily, snapshot 2026-11-13.