GPT-6 Astra Hits 80% Accuracy on IKEA Furniture Assembly Error Benchmark

OpenAI's GPT-6 Astra can look at a photo of an IKEA furniture piece and identify whether it was assembled incorrectly. The model reached 80% accuracy on Epoch AI's Furniture Assembly Benchmark, which tests models against photos of three IKEA pieces assembled with deliberate errors. The task requires comparing photos with instructions, finding mistakes, and explaining what went wrong.

The reported jump is large: in November 2025, the best model on the benchmark, Claude Opus 4.5, scored 28%. Ten months later, GPT-6 Astra reached 80%, while Claude Fable 5.1 scored 70% and Claude Opus 5 scored 61%. The Decoder also reports that Chinese open-weight models such as Kimi K3 trail the leaders by at least seven months on this benchmark.

Epoch AI says the speed is not yet fast enough for real-time assembly guidance, with GPT-6 Astra taking three minutes per photo. That limitation matters for practical use: a helpful assembly assistant would need to catch errors quickly while a person is still building. Even so, the gap is closing, and the same type of capability could eventually support tasks such as car repairs or appliance fixes.

The result points to rapid progress in visual reasoning, especially for models that need to connect images, instructions, and physical-world structure. The Decoder notes that models were failing much simpler visual tasks not long ago, making Astra's performance notable. For readers, the key takeaway is clear: photo-based troubleshooting is moving from novelty toward practical assistance, but latency remains a major hurdle.