OpenAI's GPT-6 Astra Scores 80% on IKEA Assembly-Error Benchmark, Up From 28% Ten Months Ago

OpenAI's GPT-6 Astra scored 80% on Epoch AI's Furniture Assembly Benchmark (FAB), a test that measures whether AI models can spot mistakes in IKEA furniture builds by comparing progress photos against assembly instructions. Ten months earlier, in November 2025, the best-performing model on the same benchmark—Claude Opus 4.5—scored just 28%.

FAB works by photographing three IKEA furniture pieces at various stages of assembly, with deliberate errors introduced during the build. Models are shown the photos alongside the official instructions and asked to identify what went wrong and describe the mistake. It's a test of fine-grained visual reasoning: matching physical object states against a sequence of diagrams, then localizing discrepancies.

According to Epoch AI, Claude Fable 5.1 placed second at 70%, followed by Claude Opus 5 at 61%. Chinese open-weight models, including Kimi K3, trail the leading closed models by at least seven months on this benchmark, according to the researchers.

GPT-6 Astra's 80% score comes with a catch: it takes roughly three minutes per photo to process, according to Epoch AI. That processing time rules out real-time use during an actual furniture build—by the time the model finishes analyzing a single image, a person could likely have already found their own mistake. But the researchers suggest the underlying capability could eventually extend to other structured visual-inspection tasks, such as diagnosing car repairs or troubleshooting appliance faults.

Epoch AI frames the score jump as notable given how recently multimodal models were failing much simpler visual tasks. A near-tripling of accuracy in ten months on a benchmark specifically designed to test detailed visual-instruction matching represents one of the sharper capability jumps tracked on any narrow visual benchmark this year. The report also notes that GPT-6 Astra performs strongly on visual robotic tasks more broadly, suggesting the gains extend beyond the furniture-specific test.

No pricing, context window, or architectural details for GPT-6 Astra were disclosed in the source material. Epoch AI's FAB benchmark itself remains a narrow, purpose-built test rather than a general capability suite, so the 80% figure should be read as a specific measure of visual-error-detection performance rather than a broad claim about model quality.

What this means

Fine-grained visual comparison—matching real-world object states against reference instructions and pinpointing exact discrepancies—has historically been a weak spot for multimodal models, which tend to describe images in broad strokes rather than catch small structural errors. A jump from 28% to 80% in under a year suggests this specific capability is improving faster than general benchmark trends would predict. If the three-minute-per-image latency comes down, the practical applications extend well beyond furniture: quality control in manufacturing, remote technical support, and any workflow where someone needs to verify that a physical process matches a documented procedure. The gap between closed frontier models and open-weight alternatives on this task—cited at seven-plus months—also signals that detailed visual reasoning remains one of the areas where proprietary labs are extending their lead rather than seeing it close.