AI can now spot your IKEA assembly mistake 80% of the time

Epoch AI, the research group that tracks how fast AI capabilities move, posted a new benchmark on X yesterday with a question anyone who owns an Allen key can relate to. "Can AI tell if you've built your IKEA furniture wrong?" @EpochAIResearch asked. Their answer: the top score went "from 28% to 80% in just 10 months."

That reads like a novelty, but it's one of the cleaner tests we've seen of whether AI can look at a photo of real physical work, compare it with the instructions, and say what went wrong. Plenty of small businesses do that job by hand every day.

What Epoch tested

The benchmark is called FAB, the Furniture Assembly Benchmark, and Aiden Ament and Greg Burnham built it at Epoch. Each model gets the IKEA manual and a photo of a piece of furniture partway through assembly. Some photos show a correct build and some show a deliberate mistake. The model has to say whether something is wrong, name the step where it went wrong, and describe the error well enough to satisfy a grader.

There are 60 photos across three builds of rising difficulty: a STÄLL shoe cabinet (about an hour and a half to assemble), a TONSTAD bed frame (about three and a half hours) and a GULLABERG dresser (about five). Models could zoom into the images and run code while they worked.

Pay attention to the scoring. A model that calls a mistake on a correct build loses points, the same as one that misses a real error. Once you try to use this at work, that distinction decides whether the tool helps or just makes noise.

The numbers

According to Epoch's writeup:

• GPT-6 Astra leads at 80%, with a median of about three minutes per photo, the fastest of the models tested.
• Claude Fable 5.1 scored 70% and Claude Opus 5 scored 61%.
• The best score in November 2025 was 28%, from Claude Opus 4.5.
• Open-weight models trail the closed frontier by at least seven months.

We found the failure patterns more useful than the leaderboard. Epoch reports that Google's Gemini and Alibaba's Qwen models "almost always assume a mistake exists." Older Anthropic and OpenAI models had the opposite habit and waved real errors through. Difficulty also tracked how subtle the mistake was rather than how complex the furniture was, so the dresser wasn't automatically harder than the shoe cabinet. A panel turned the wrong way round was.

Our read

A jump from 28% to 80% in ten months is big, and we take it at face value. The authors are careful about the limits, though. Sixty photos from three builds is a small sample, and they say plainly that it's unclear how well the results carry over to other physical tasks. Three minutes per photo works for a review queue. It's too slow for someone standing in a basement who needs an answer now.

The fair takeaway is narrower than "AI can inspect work." AI has become a decent second set of eyes on a photo, as long as you hand it the reference document and a person makes the final call. Epoch's two failure modes also tell you what to test before you rely on any of this. A model that flags everything wastes your crew's time, and a model that flags nothing gives you false comfort. You want to know which one you have.

What this means for a business that runs on photos

A lot of trades already ask techs to snap a picture when they finish a job: a furnace install, a water heater hookup, an electrical panel, a finished bathroom. Mostly those photos sit in a folder until a customer complains. The FAB results point to a cheap middle step. Compare each photo against the spec or checklist for that job, flag anything that looks off, and have a lead review only the flagged ones.

Think of it as triage for your install photos. It only works if the reference material is written down (manuals, install specs, your own checklists) and someone owns the job of reviewing flags quickly. In our experience, getting that paperwork in order takes more effort than wiring up the AI.

If you run a Fox Valley business and want to know where a photo check or any other AI step would save real hours, New Face Design offers a free process audit. We map how a job moves through your shop today and tell you which parts are ready for automation and which still need a person.