GPT-6 Astra Tops IKEA Assembly Test at 28%

Epoch AI runs a test called the Furniture Assembly Benchmark, or FAB, and it does exactly what the name suggests: it photographs three IKEA pieces mid-build, some with deliberate errors baked in, and asks AI models to compare the photos against the instructions and spot what went wrong. Last November, the best-performing model, Claude Opus 4.5, managed a score of 28%. That's not a typo. A quarter century of household frustration with Swedish flat-pack furniture and the leading AI model could barely tell you if you'd installed a shelf pin backwards.

Ten months is apparently long enough to fix that. OpenAI's GPT-6 Astra now scores 80% on the same benchmark, though it takes about three minutes to process each photo. Anthropic's Claude Fable 5.1 sits at 70%, with Claude Opus 5 at 61%. Chinese open-weight models, including Kimi K3, are running roughly seven months behind the frontier.

The jump matters less for shelf-building and more for what it signals: models that recently struggled with basic visual reasoning are now catching detailed assembly errors from a photo. Researchers see the same underlying capability eventually stretching toward car repairs or appliance troubleshooting, once the speed catches up to the accuracy.