GPT-6 Astra identifies IKEA assembly errors with 80% accuracy
OpenAI’s GPT-6 Astra can identify incorrect assembly in photographs of IKEA furniture, according to the Furniture Assembly Benchmark (FAB) from Epoch AI. The benchmark uses images of three IKEA pieces being assembled with intentional mistakes. Models compare each image with the relevant instructions, locate the errors, and explain what went wrong.
In November 2025, the leading system, Claude Opus 4.5, achieved 28% accuracy. Ten months later, GPT-6 Astra reached 80%, taking three minutes to process each photograph. Claude Fable 5.1 scored 70%, while Claude Opus 5 reached 61%. Chinese open-weight models, including Kimi K3, were at least seven months behind the leaders, according to the report.
The system remains too slow to provide real-time guidance during assembly. Researchers said the technology could eventually be useful for diagnosing problems with car repairs or household appliances. They also highlighted Astra’s performance on visual robotic tasks and the broader progress from models that recently struggled with simpler visual challenges.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.