AI now can spot furniture assembly errors with 80% accuracy using instructions and photos

On September 26, 2026, Epoch AI announced results from the Furniture Assembly Benchmark (FAB), a test designed to evaluate AI's ability to understand real-world assembly processes. OpenAI's GPT-6 Astra achieved an 80% accuracy rate in identifying errors in IKEA furniture assembly, a significant improvement from 28% in November 2025. The test involved 60 photos of incorrectly assembled furniture, with AI models required to compare images with official manuals to detect and explain mistakes. The test provided AI with access to PDF manuals, image zoom tools, and Python interpreters, simulating real-world scenarios where users might take photos of their assembly progress. GPT-6 Astra completed single-photo analysis in about 3 minutes, the fastest among tested models. Researchers noted that models like Gemini and Qwen often assumed errors existed before searching, while earlier models like Claude Opus 4.5 struggled to find any. In China, the best open-source model, Kimi K3, lagged behind international models by about seven months, highlighting a gap in visual reasoning capabilities. DeepSeek V4 Pro and GLM 5.3, however, did not provide image support, limiting their performance in this test.