OCOPD GPT-6 Astra has sharply improved the ability of AI systems to identify assembly errors in physical objects, scoring 80% on Epoch AI's Furniture Assembly Benchmark. The result compares with a leading score of 28% in November 2025, when Anthropic's Claude Opus 4.5 topped the evaluation.
The benchmark tests whether a vision-capable model can inspect a photograph of partially assembled IKEA furniture, compare it with technical instructions, determine whether the construction is correct, and locate the mistake. The latest results suggest that multimodal AI is becoming substantially more capable at reasoning about the physical world, not just describing images or answering questions about them.
A rapid gain in visual construction reasoning
Epoch AI evaluated models using 60 photographs of half-built IKEA furniture containing intentional mistakes. GPT-6 Astra achieved the highest reported score, followed by Claude Fable 5.1 at 70% and Claude Opus 5 at 61%.
The improvement is notable both in absolute accuracy and in the pace of progress:
- The top score increased from 28% to 80% in ten months.
- GPT-6 Astra analyzed an image in a median of three minutes.
- That response time was between two and ten times faster than previously leading models evaluated by Epoch AI.
- The best Chinese open-weight model reportedly remained at least seven months behind the frontier.
The benchmark did not find a clear relationship between an item's IKEA Complexity Index and model performance. Epoch AI instead pointed to factors such as the subtlety of the error, the camera angle, and how far construction continued after the mistake. Those variables are important because a practical assistant must recognize errors that may be partially hidden or become harder to diagnose as more pieces are added.
Accuracy is improving faster than usability
An 80% score marks meaningful progress, but it does not yet make GPT-6 Astra a dependable replacement for human inspection. Missing one in five cases could still lead users to continue building on an incorrect foundation, particularly when the model's explanation is treated as authoritative.
Speed is another constraint. A median analysis time of three minutes represents a major improvement, but it is still too slow for smooth, real-time guidance during assembly. Users typically need feedback while positioning a part or tightening a fastener, rather than several minutes after taking a photograph.
The results also expose two broad failure patterns among evaluated systems. Some models tended to over-flag potential errors, while others more often missed genuine mistakes. Both behaviors reduce trust: excessive warnings create unnecessary rework, while missed errors can compound into larger construction problems.
Implications for AI assistants and device makers
The benchmark points toward applications beyond IKEA furniture. Models that can compare a real-world object with diagrams and identify deviations could support equipment inspection, repair, manufacturing quality control, and guided maintenance. The same capability could eventually be integrated into smartphone cameras, augmented-reality tools, customer-support systems, or robotic workflows.
For AI developers, the findings also highlight a gap in conventional vision evaluations. Image captioning, visual question answering, and document parsing do not adequately measure whether a model can reason about spatial relationships in a physical object and connect those observations to procedural instructions. Benchmarks such as FAB may therefore become more relevant as companies compete to demonstrate useful embodied intelligence.
The competitive picture remains fluid. GPT-6 Astra's lead shows that frontier models can make large gains in a relatively short period, while the performance gap between closed-weight leaders and open-weight systems remains substantial in this test. The immediate commercial impact is likely to be strongest in assisted inspection and troubleshooting, where users can tolerate some delay and verify recommendations themselves. Fully autonomous or real-time physical guidance will require further improvements in accuracy, latency, reliability, and the transparency of model decisions.
Source
Original source: the-decoder.com