OpenAI's GPT-6 Astra can now tell you exactly where you screwed up your IKEA shelf
decoder ·
OpenAI's GPT-6 Astra model has achieved an 80% accuracy rate in identifying errors during IKEA furniture assembly, a significant improvement from the previous best score of 28%. This performance was measured using Epoch AI's Furniture Assembly Benchmark (FAB), which tests models by comparing photos of deliberately flawed builds against official instructions. While GPT-6 Astra leads with its high accuracy, other models like Claude Fable 5.1 and Claude Opus 5 also demonstrated strong capabilities at 70% and 61% respectively. Chinese open-weight models such as Kimi K3 currently lag behind these leaders by several months. Although the current processing speed of three minutes per photo is too slow for real-time assistance, researchers believe the technology holds promise for future applications in car repairs and appliance fixes. The rapid progress highlights how quickly AI models have advanced in handling complex visual tasks that previously challenged them. This development underscores the growing capability of large language models to interpret physical world interactions through visual data.
decoder