OpenAI’s GPT-6 Astra has achieved an 80% accuracy rate in identifying assembly errors on three specific IKEA furniture pieces, a significant jump from the 28% scored by the best previous model in November 2025. Epoch AI researchers tested this capability using their Furniture Assembly Benchmark, which photographs deliberate mistakes during construction and compares them against official instructions. While Claude Fable 5.1 and Claude Opus 5 followed at 70% and 61% respectively, the new Astra model completes this analysis in just three minutes per photo. Chinese open-weight models such as Kimi K3 remain at least seven months behind these leaders.
This performance marks a notable shift from earlier failures on simpler visual tasks, suggesting the technology is maturing beyond basic description. Although the current speed is too slow for immediate real-time guidance, the underlying capability could eventually assist with complex repairs for cars or household appliances. The system also demonstrates strong performance in visual robotic tasks, indicating broader utility for physical world interactions.
- Analysis time reduced to three minutes per photo
- Accuracy improved from 28% to 80% in ten months
- Applicable to car repairs and appliance fixes
