Furniture Assembly Benchmark
Multimodal · 2026-09-23
Epoch's 60 photographed IKEA assemblies across three products. Given the photo and assembly manual, an agent identifies whether a mistake exists and its step and description, using zoom and Python with an 80-step budget. Reports accuracy, with a model judge checking error descriptions.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 83.3 |
| GPT-6.1 Sol | 80.0 |
| GPT-6 Astra | 80.0 |
| Claude Sonnet 5.5 | 75.0 |
| Gemini 4 Argon | 72.5 |
| Claude Fable 5.1 | 70.0 |
| Claude Opus 5 | 60.8 |
| GPT-6 Sol | 58.3 |
| GPT-5.6 Sol | 56.7 |
| GPT-5.6 Terra | 54.2 |
| Muse Wonder | 54.2 |
| Claude Haiku 5.5 | 47.5 |
| GPT-5.5 | 44.2 |
| GPT-6 Luna | 44.2 |
| Opus 4.8 | 42.5 |
| GPT-5.6 Luna | 42.5 |
| Grok 4.6 | 40.0 |
| GPT-5.2 | 38.3 |
| GPT-5.4 | 37.5 |
| Claude Fable 5 | 35.8 |
| DeepSeek V4.1 Flash | 34.2 |
| Kimi K3 | 34.2 |
| Opus 4.7 | 33.3 |
| Gemini 3.8 Flash | 31.7 |
| Opus 4.5 | 28.3 |
| Opus 4.6 | 28.3 |
| Gemini 3.1 Pro Preview | 26.7 |
| Gemini 3.7 Flash | 26.7 |
| Gemini 3.6 Flash | 23.3 |
| Grok 4.5 | 22.5 |
| Kimi K2.6 | 21.7 |
| Grok 4.7 | 20.8 |
| Mistral Large 4 Preview | 20.8 |
| Qwen3.8 Max (0902) | 20.0 |