GPT-6 Astra Spots IKEA Furniture Assembly Mistakes From Photos

Epoch AI presented a challenge to their AI models. Given an image of completed IKEA furniture, identify where the user made a mistake. At the time of writing (November 2025), the leading AI model for this challenge, Claude Opus 4.5, is able to accurately detect mistakes in 28% of cases. Fast forward ten months, and the latest AI model from OpenAI, GPT-6 Astra, can accurately detect mistakes in 80% of cases. However, GPT-6 Astra takes approximately 3 minutes to analyze each image.

The three minute analysis time is far more important than the accuracy metric. If this model takes so long to process each image, then it isn’t useful for providing real time feedback during assembly. Instead, it’s useful for postmortem analysis, when you already have some furniture assembled incorrectly and you need help figuring out what went wrong. To me, the interesting part here isn’t the accuracy rate, but rather the fact that the AI was able to learn how to compare an image to written directions and pinpoint the correct location within the text. Soon enough, we’ll have a quick and affordable way to perform this type of analysis.