AI Reports · Epoch AI · 2026-09-23 · Blog / commentary
Can AI spot mistakes in IKEA assembly?
Epoch AI's Furniture Assembly Benchmark (FAB) tests visual and spatial reasoning with 60 photos from three IKEA furniture builds, some correctly assembled and some with intentional mistakes. The best score rose from 28% in November 2025 (Claude Opus 4.5) to 80% for GPT-6 Astra, which has a median time of 3 minutes per photo. Chinese open-weight models lag the frontier by at least seven months.