GPT-6 Astra Flags Bad IKEA Builds at 80% Accuracy, Up From 28% in November

OpenAI's GPT-6 Astra can now look at a photograph of an assembled piece of IKEA furniture and identify whether it was put together wrong, scoring 80 percent accuracy on the task. That is a steep climb from November 2025, when the strongest model available managed only 28 percent. The result, flagged by The Decoder, puts a concrete number on how much better the new model is at reading physical objects and their mistakes.

The jump matters because spotting assembly errors is harder than it sounds. A model has to recognize a shelf or cabinet, work out which panels and fittings are meant to meet where, and then notice when a board is flipped, a hole is on the wrong side or a piece is attached in the wrong orientation. Those are spatial and procedural judgments, not text prediction, and the November baseline of 28 percent showed how far short earlier systems fell.

Epoch AI, which tracks model capabilities, weighed in on the result. Its assessment is that the speed is not yet high enough for real-time assembly guidance, meaning Astra cannot watch you build in the moment and call out a mistake as you make it. The accuracy is there for a still photo, but latency remains the bottleneck before this becomes a live helper.

The gap between 28 percent and 80 percent in roughly half a year follows a pattern in vision-language work, where each model generation tends to lift the ceiling on tasks that were previously near-unusable. What is different here is the target: a physical, hands-on chore with a clear right and wrong answer, rather than a benchmark built from text or curated images.

For OpenAI, the demo doubles as a signal that GPT-6 is being positioned around practical visual reasoning, not just conversation. A model that can audit a half-built bookcase is also a model that can inspect a wiring harness, a packaged part or a finished assembly line item, and those are markets where buyers care about accuracy and response time in equal measure.

The open question is latency. If Astra's turnaround time keeps falling at the rate its accuracy did, real-time coaching for assembly, repair and inspection stops being a demo and becomes a product feature. Until then, the model can tell you where you went wrong after the fact, which is useful but a step behind the moment that costs you an hour with an Allen key.