OpenAI GPT-6 Astra Scores 80% on IKEA Assembly Test

OpenAI takes a step forward in visual testing: GPT-6 Astra achieves 80% accuracy in spotting assembly errors in IKEA furniture, with three minutes of analysis per photo. The model outperforms the latest versions from Anthropic, while still being too slow for real-time assistance. Researchers are considering future applications for automotive or appliance repairs.

No real-time assistance and potential technical uses

The technology is not yet fast enough to guide assembly in real-time. However, researchers believe it could eventually be applied to repair tasks on cars or household appliances. Astra also excels in visual robotic tasks. The progress made, particularly with Astra, is considered remarkable given the recent challenges faced with simpler visual tasks.

Scores: 80% for Astra, ahead of Anthropic models

Ten months after the previous benchmark, GPT-6 Astra achieves 80% accuracy with a time of three minutes per photo. Anthropic's models follow: 70% for Claude Fable 5.1 and 61% for Claude Opus 5. In November 2025, the best score was 28% with Claude Opus 4.5. The observed progression shows an increase from 28% to 80% in just a few months.

What FAB measures and the status of open-weight models

The Furniture Assembly Benchmark from Epoch AI involves photographing three IKEA pieces during assembly, introducing deliberate errors. These models analyze the obtained images by comparing them to the instructions, identifying mistakes, and describing the source of the problem. Some open-weight Chinese models, such as Kimi K3, show a lag of at least seven months behind the most advanced models.