A-Fantasia Cube Rotation
General QA · 2025-04-25
A-Fantasia Cube tests spatial tracking by assigning six distinct colors to cube faces, applying one to three random quarter-turn rotations, and asking for the color on a target face. Scores are error rates over 100 trials, so lower is better.
Top models (lower is better)
| Model | Score |
|---|---|
| Opus 4.6 | 53.0 |
| Gemini 3 Flash Preview | 57.0 |
| Gemini 3.1 Flash-Lite | 60.0 |
| Qwen3 Max (rolling alias) | 62.0 |
| GPT-5.5 | 63.0 |
| GPT-5.6 Sol | 63.0 |
| Gemini 1.5 Pro | 64.0 |
| Gemini 3.1 Pro Preview | 66.0 |
| Qwen3.7-Plus | 66.0 |
| Gemini 1.5 Flash | 66.0 |
| Opus 4.5 | 68.0 |
| Gemini 3 Pro Preview | 68.0 |
| GPT-5.4 | 68.0 |
| Claude 3.7 Sonnet | 68.0 |
| Gemini 2.0 Flash 001 | 68.0 |
| GLM-5.1 | 68.0 |
| Llama 3.1 405B Instruct | 68.0 |
| DeepSeek-V3.2-Exp | 68.0 |
| Claude 3.5 Sonnet (Oct 2024) | 69.0 |
| Qwen3.6 Plus (2026-04-02) | 69.0 |
| Opus 4 | 70.0 |
| GLM-5 | 70.0 |
| DeepSeek-V3-0324 | 70.0 |
| Sonnet 4 | 71.0 |
| DeepSeek-V3.1 | 71.0 |
| Opus 4.1 | 72.0 |
| Sonnet 4.5 | 72.0 |
| GLM-5.2 | 72.0 |
| MiniMax M3 | 72.0 |
| GPT-5.6 Luna | 73.0 |
| Nemotron 3 Ultra 550B A55B | 73.0 |
| GPT-4o | 74.0 |
| Opus 3 | 74.0 |
| GPT-5.1 | 74.0 |
| Grok 4.20 | 74.0 |
| GPT-5.6 Terra | 75.0 |
| GPT-5.2 | 75.0 |
| Gemini 2.5 Flash | 75.0 |
| Llama 3.3 70B Instruct | 75.0 |
| Mistral Large 2.1 (Instruct 2411) | 75.0 |