Epoch AI Furniture Assembly Test: GPT-6 Astra Hits 80% Accuracy as Multimodal Visual Reasoning Improves

Epoch AI released results from its Furniture Assembly Benchmark (FAB) on September 23 local time, designed to evaluate how well multimodal AI understands real-world assembly processes. The test used 60 photos of IKEA furniture assembly and asked AI to compare them against official manuals, identify intentionally placed errors, and explain the specific problems.

OpenAI's GPT-6 Astra topped the list with 80% accuracy, while Anthropic's Claude Fable 5.1 and Claude Opus 5 reached 70% and 61% respectively. Compared to the leading model Claude Opus 4.5 in November 2025, which scored only 28%, frontier models have improved significantly in about 10 months. GPT-6 Astra took a median of about 3 minutes per photo analysis, making it the fastest model in the study.

The research noted different error patterns across models: Google Gemini and Alibaba Qwen series tend to assume errors exist before searching for them, while earlier Anthropic and OpenAI models often failed to find errors at all. Among Chinese models, the best open-weight model Kimi K3 lags about 7 months behind international frontier models, with visual reasoning remaining a relatively weak area for some Chinese models.