AI detects errors in the assembly of IKEA furniture

In the Furniture Assembly Benchmark by Epoch AI, GPT-6 Astra identifies 80 percent of the errors in IKEA furniture. For a real-time inspection, the model is still too slow according to the researchers.

A photo of a piece of furniture is now enough for an AI model to recognize a faulty assembly step. In the Furniture Assembly Benchmark by Epoch AI, models had to match photos with an assembly instruction and describe errors in the setup. The results show how the systems perform in this specific inspection task – and where their limits are.

GPT-6 Astra recognizes 80 percent of the errors

For the benchmark, the researchers photographed three IKEA furniture pieces during assembly and deliberately introduced errors. The models were to find and correctly describe faulty steps based on the photos and the instruction.

In November 2025, Anthropics Claude Opus 4.5 recognized 28 percent of the errors according to Epoch AI. Ten months later, OpenAI's GPT-6 Astra achieved 80 percent. Other results from the comparison: Claude Fable 5.1 reached 70 percent, Claude Opus 5 came in at 61 percent. Chinese open-weight models are reportedly at least seven months behind the leading models according to Epoch AI.

Still no real-time inspection

According to the report, GPT-6 Astra needed three minutes per photo. While it was the fastest model in the comparison, researchers believe that the speed is still not sufficient for real-time use in furniture assembly. The results pertain to the described benchmark; they do not prove that a system is already available as a practical assembly aid.

What this means for practice

For you as a designer or in another profession, the approach is particularly interesting: a model matches image material with an instruction and looks for discrepancies. This could also become relevant for other practical inspection tasks in the future. The source mentions possible application areas such as auto repairs and household appliances.

For the moment, however, you should categorize the results as the state of research on a limited test task. The detection rate of GPT-6 Astra is significantly higher than that of Claude Opus 4.5 in the earlier comparison. The necessary time per photo and the gap to a real-time inspection also indicate that such an application is not yet easily transferable to everyday work.

Sources

• THE DECODER: AI models learn to read IKEA assembly instructions and find errors in assembly