GPT-6 Astra spots IKEA assembly mistakes at 80 percent, up from 28

OpenAI's GPT-6 Astra now finds most of the mistakes people make while assembling IKEA furniture. It scores 80 percent on Epoch AI's Furniture Assembly Benchmark (FAB), which the research group published on September 23. In November 2025 the best result was 28 percent, set by Anthropic's Claude Opus 4.5.

The benchmark asks a narrow but practical question: can a model connect a paper instruction manual to a real object and explain what went wrong? Epoch AI treats it as a proxy for visual and spatial reasoning.

What the benchmark tests

Epoch AI bought three IKEA pieces, chosen by the IKEA Complexity Index: the STÄLL shoe rack, the TONSTAD bed frame and the GULLABERG dresser. Researchers photographed them during assembly and deliberately made realistic errors, often building past a mistake to make it harder to spot.

Each model sees 60 photos, some of correct builds and some with mistakes, plus the manual, a zoom tool and a Python interpreter inside an 80-step agentic budget. To score, a model must name every step that contains a mistake and describe it; a lenient LLM grader, GPT-5.6 Sol, then checks the description.

Rivals and the open-weight gap

Claude Fable 5.1 follows at 70 percent and Claude Opus 5 at 61 percent. Chinese open-weight models trail the frontier: Kimi K3 lags by seven months on FAB, against 4.4 months on Epoch's general Capabilities Index. Epoch AI notes that image support is still rare in Chinese models, with DeepSeek V4 Pro, DeepSeek Flash and Z.ai's GLM 5.3 all lacking it.

Why it matters

Furniture assembly stands in for maintenance work that needs the same skills, such as repairing a car or a household appliance, the researchers write. Astra is also the fastest model tested, at a median of three minutes per photo, two to ten times quicker than earlier leaders, though still too slow for real-time guidance. The test covers only 60 photos from three builds, so how far the results generalise is unclear.