Benchmarks › Furniture Assembly

Furniture Assembly

Inspect furniture photos and manuals to identify assembly mistakes and their steps
83.3top score · Claude Opus 5.5
31models ranked
2open-weights models
Accuracy across 60 photoshigher is better
Oct 20, 2026last updated

How the test works

1

Task

60 assembly photos

  • What the model doesInspect furniture photos and manuals to identify assembly mistakes and their steps
  • Kinds of tasksAssembly error detection, spatial reasoning
2

Run

Agentic sandbox

  • Step limit80 steps
  • Tools allowedImage zoom, Python interpreter
3

Score

Accuracy across 60 photos

  • Graded byCorrect build decision; for errors, correct steps plus lenient GPT-5.6 Sol description grading
  • Better isHigher
Sources: epoch.ai

Furniture Assembly leaderboardtop 10 of 31

Full leaderboard with filters →
ModelFurniture AssemblyPer taskPriceSpeedSupportsContext
83.3max
$5.98$4 / $20961M
80max
$3.26$10 / $50691.05M
80max
$0.72$2 / $10521.05M
75max
$5.46$2 / $101371M
72.5high
$1.99$2 / $10841M
70max
$7.63$10 / $50691M
Claude Opus 5Anthropic
60.8max
$5.86$5 / $25551M
GPT-6 SolOpenAI
58.3max
$1.04$2 / $101281.05M
56.7max
$1.99$4 / $201141.05M
54.2max
—$4 / $20—1M

Furniture Assembly questions

What is Furniture Assembly?›

Inspect furniture photos and manuals to identify assembly mistakes and their steps. It has 60 assembly photos. Created by Epoch AI.

Which model is best on Furniture Assembly?›

Claude Opus 5.5 leads Furniture Assembly with 83.3, ahead of GPT-6 Astra (80) and GPT-6.1 Sol (80).

What is the best open-weights model on Furniture Assembly?›

Kimi K3 with 34.2, then Kimi K2.6 (21.7).

How is Furniture Assembly scored?›

Accuracy across 60 photos; correct build decision; for errors, correct steps plus lenient gpt-5.6 sol description grading.