OpenAI’s GPT-6 Astra achieved an 80 percent accuracy score on Epoch AI’s Furniture Assembly Benchmark (FAB), identifying structural assembly errors in IKEA furniture from photographs. The benchmark tests models by asking them to compare pictures of partially assembled items containing deliberate mistakes against instruction manuals.

The score marks a significant jump in visual spatial reasoning capabilities compared to previous benchmark results. In November 2025, the top-performing model, Claude Opus 4.5, scored 28 percent. GPT-6 Astra outperformed competing models including Claude Fable 5.1 at 70 percent and Claude Opus 5 at 61 percent.

While processing times of three minutes per photo remain too slow for real-time consumer guidance, researchers note that Astra’s capabilities demonstrate rapid progress in visual perception. Potential future applications include assisting with physical appliance maintenance, car repairs, and robotic spatial tasks.

Why it matters

  • Multimodal visual reasoning is advancing rapidly from basic image recognition to multi-step 3D physical error detection.

  • High processing latency remains a critical bottleneck for deploying visual reasoning models in real-time physical AI applications.

Source: the-decoder.com