Benchmarks / Reasoning
Furniture Assembly
Epoch AI's spatial-reasoning benchmark: from photos of IKEA builds (Ställ shoe cabinet, Tonstad bed, Gullaberg dresser) and the official manual, the model decides whether the build is correct and, if not, identifies and describes the mistaken step. Mean score. Data: Epoch AI, CC BY.
Furniture Assembly
Higher is better · best effort per model
full table below
| # | Model | Effort | Score |
|---|---|---|---|
| 1 | Opus 5.5 | max | 83.3% |
| 2 | GPT-6.1 Sol | max | 80.0% |
| 3 | GPT-6 Astra | max | 80.0% |
| 4 | Sonnet 5.5 | max | 75.0% |
| 5 | Gemini 4 | high | 72.5% |
| 6 | Fable 5.1 | max | 70.0% |
| 7 | Opus 5 | max | 60.8% |
| 8 | GPT-6 Sol | max | 58.3% |
| 9 | GPT-5.6 Sol | max | 56.7% |
| 10 | GPT-5.6 Terra | max | 54.2% |
| 11 | Haiku 5.5 | max | 47.5% |
| 12 | GPT-6 Luna | max | 44.2% |
| 13 | GPT-5.5 | xhigh | 44.2% |
| 14 | Opus 4.8 | max | 42.5% |
| 15 | GPT-5.6 Luna | max | 42.5% |
| 16 | Grok 4.6 | xhigh | 40.0% |
| 17 | GPT-5.2 | xhigh | 38.3% |
| 18 | GPT-5.4 | xhigh | 37.5% |
| 19 | Fable 5 | max | 35.8% |
| 20 | DeepSeek V4.1 Flash | max | 34.2% |
| 21 | Kimi K3 | max | 34.2% |
| 22 | Opus 4.7 | max | 33.3% |
| 23 | Gemini 3.8 Flash | high | 31.7% |
| 24 | Opus 4.6 | max | 28.3% |
| 25 | Opus 4.5 | high | 28.3% |
| 26 | Gemini 3.7 Flash | high | 26.7% |
| 27 | Gemini 3.1 Pro | high | 26.7% |
| 28 | Gemini 3.6 Flash | high | 23.3% |
| 29 | Grok 4.5 | high | 22.5% |
| 30 | Kimi K2.6 | max | 21.7% |
| 31 | Mistral Large 4 | high | 20.8% |
| 32 | Grok 4.7 | xhigh | 20.8% |
| 33 | Qwen 3.8 Max | xhigh | 20.0% |