# Furniture Assembly Benchmark (Epoch AI) (Furniture Assembly)

> Epoch AI's Furniture Assembly benchmark shows models photos of IKEA furniture builds with the official manual and asks whether the build is correct so far and, if not, which step went wrong.

Canonical page: https://benchlm.ai/benchmarks/furnitureassembly

- Category: [Multimodal & Grounded](/multimodal-grounded)
- Last updated: November 13, 2026

## About Furniture Assembly

- Year: 2026
- Tasks: 60 photos of three IKEA builds, some with deliberate assembly mistakes
- Format: Agentic visual check with manual PDF, zoom tool and Python, 80-step budget
- Difficulty: Spatial reasoning from technical diagrams to real photos
- Paper: [Furniture Assembly Benchmark (Epoch AI)](https://epoch.ai/benchmarks/furniture-assembly)

Epoch built three IKEA products of increasing complexity (the Ställ shoe cabinet, Tonstad bed frame and Gullaberg dresser) and took 60 photos during assembly. Some show correct builds; others contain a deliberate, clearly visible mistake, often with assembly continued past it. For each photo the model gets the product, the step just completed, the photo, the assembly manual as a PDF, a zoom tool and a Python interpreter, with a budget of 80 agent steps. A correct build must be called correct; for a faulty build the model must name the mistaken step, and an LLM grader then checks the description leniently.

Furniture Assembly is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (37 models)

| Rank | Model | Creator | Score | Who ran it |
|------|-------|---------|-------|-----------|
| 1 | [Claude Fable 5.2](/models/claude-fable-5-2) | Anthropic | 90.0% | Run by Furniture Assembly |
| 2 | [Claude Opus 5.5](/models/claude-opus-5-5) | Anthropic | 83.3% | Run by Furniture Assembly |
| 3 | [GPT-6 Astra](/models/gpt-6-astra) | OpenAI | 80.0% | Run by Furniture Assembly |
| 4 | [GPT-6.1 Sol](/models/gpt-6-1-sol) | OpenAI | 80.0% | Run by Furniture Assembly |
| 5 | [GPT-6.1 Terra](/models/gpt-6-1-terra) | OpenAI | 79.2% | Run by Furniture Assembly |
| 6 | [Gemini 4 Carbon](/models/gemini-4-carbon) | Google | 77.5% | Run by Furniture Assembly |
| 7 | [Claude Sonnet 5.5](/models/claude-sonnet-5-5) | Anthropic | 75.0% | Run by Furniture Assembly |
| 8 | [Gemini 4 Argon](/models/gemini-4-argon) | Google | 72.5% | Run by Furniture Assembly |
| 9 | [Claude Fable 5.1](/models/claude-fable-5-1) | Anthropic | 70.0% | Run by Furniture Assembly |
| 10 | [Claude Opus 5](/models/claude-opus-5) | Anthropic | 60.8% | Run by Furniture Assembly |
| 11 | [GPT-6 Sol](/models/gpt-6-sol) | OpenAI | 58.3% | Run by Furniture Assembly |
| 12 | [GPT-5.6 Sol](/models/gpt-5-6-sol) | OpenAI | 56.7% | Run by Furniture Assembly |
| 13 | [Kimi K3.1 Pro](/models/kimi-3-1-pro) | Moonshot AI | 55.8% | Run by Furniture Assembly |
| 14 | [GPT-5.6 Terra](/models/gpt-5-6-terra) | OpenAI | 54.2% | Run by Furniture Assembly |
| 15 | [Muse Wonder](/models/muse-wonder) | Meta | 54.2% | Run by Furniture Assembly |
| 16 | [Claude Haiku 5.5](/models/claude-haiku-5-5) | Anthropic | 47.5% | Run by Furniture Assembly |
| 17 | [GPT-5.5](/models/gpt-5-5) | OpenAI | 44.2% | Run by Furniture Assembly |
| 18 | [GPT-6 Luna](/models/gpt-6-luna) | OpenAI | 44.2% | Run by Furniture Assembly |
| 19 | [Claude Opus 4.8](/models/claude-opus-4-8) | Anthropic | 42.5% | Run by Furniture Assembly |
| 20 | [GPT-5.6 Luna](/models/gpt-5-6-luna) | OpenAI | 42.5% | Run by Furniture Assembly |
| 21 | [Grok 4.6](/models/grok-4-6) | xAI | 40.0% | Run by Furniture Assembly |
| 22 | [GPT-5.2](/models/gpt-5-2) | OpenAI | 38.3% | Run by Furniture Assembly |
| 23 | [GPT-5.4](/models/gpt-5-4) | OpenAI | 37.5% | Run by Furniture Assembly |
| 24 | [Claude Fable 5](/models/claude-fable) | Anthropic | 35.8% | Run by Furniture Assembly |
| 25 | [Kimi K3](/models/kimi-k3) | Moonshot AI | 34.2% | Run by Furniture Assembly |
| 26 | [DeepSeek V4.1 Flash 1017](/models/deepseek-v4-1-flash) | DeepSeek | 34.2% | Run by Furniture Assembly |
| 27 | [Claude Opus 4.7](/models/claude-opus-4-7) | Anthropic | 33.3% | Run by Furniture Assembly |
| 28 | [Gemini 3.8 Flash](/models/gemini-3-8-flash) | Google | 31.7% | Run by Furniture Assembly |
| 29 | [Claude Opus 4.6](/models/claude-opus-4-6) | Anthropic | 28.3% | Run by Furniture Assembly |
| 30 | [Claude Opus 4.5](/models/claude-opus-4-5) | Anthropic | 28.3% | Run by Furniture Assembly |
| 31 | [Gemini 3.7 Flash](/models/gemini-3-7-flash) | Google | 26.7% | Run by Furniture Assembly |
| 32 | [Gemini 3.1 Pro](/models/gemini-3-1-pro) | Google | 26.7% | Run by Furniture Assembly |
| 33 | [Gemini 3.6 Flash](/models/gemini-3-6-flash) | Google | 23.3% | Run by Furniture Assembly |
| 34 | [Grok 4.5](/models/grok-4-5) | xAI | 22.5% | Run by Furniture Assembly |
| 35 | [Kimi K2.6](/models/kimi-2-6) | Moonshot AI | 21.7% | Run by Furniture Assembly |
| 36 | [Mistral Large 4](/models/mistral-large-4) | Mistral | 20.8% | Run by Furniture Assembly |
| 37 | [Grok 4.7](/models/grok-4-7) | xAI | 20.8% | Run by Furniture Assembly |

## FAQ

### What does the Furniture Assembly benchmark measure?

Spatial reasoning about physical construction: reading an IKEA manual, comparing it with a photo of a real build and spotting whether, and at which step, something was assembled wrong.

### Which model is best at Furniture Assembly?

Claude Fable 5.2 at max effort leads with 90%, followed by Claude Opus 5.5 (83.3%), GPT-6.1 Sol and GPT-6 Astra (80% each).

## Compare Top Models on Furniture Assembly

- [Claude Fable 5.2 vs Claude Opus 5.5](/compare/claude-fable-5-2-vs-claude-opus-5-5)
- [Claude Opus 5.5 vs GPT-6 Astra](/compare/claude-opus-5-5-vs-gpt-6-astra)
- [GPT-6 Astra vs GPT-6.1 Sol](/compare/gpt-6-1-sol-vs-gpt-6-astra)
- [GPT-6.1 Sol vs GPT-6.1 Terra](/compare/gpt-6-1-sol-vs-gpt-6-1-terra)