Epoch AI benchmark shows AI progress in detecting furniture assembly errors

A new study by the research organization Epoch AI has demonstrated significant advancements in artificial intelligence's ability to identify errors in furniture assembly. Using the Furniture Assembly Benchmark (FAB), researchers tested AI models on their ability to detect mistakes in the assembly of three specific IKEA products: the STÄLL shoe rack, the TONSTAD bed frame, and the GULLABERG chest of drawers.

The benchmark provides models with assembly photos and the original instruction manual, requiring them to identify and describe errors. Results showed a dramatic improvement in performance over a ten-month period. While Anthropic’s Claude Opus 4.5 achieved only a 28 percent success rate in November 2025, OpenAI’s GPT-6 Astra reached an 80 percent success rate as of September 23, 2026, while also maintaining high processing speeds.

Other high-performing models included Claude Fable 5.1 with 70 percent and Claude Opus 5 with 61 percent. The testing process utilized a secondary model, GPT-5.6 Sol, to act as an automated judge to verify the accuracy of the identified errors and descriptions.