You ask AI for one small fix. It breaks the layout.
Figma says that when it started, "frontier models couldn’t understand Figma syntax", so it trained its own editing model, graded by "expert-written criteria". On 445 real-world design edits it scores 77.8% to GPT 6.1 Sol's 76.9%, and since beta "users undo the agent’s work 60% less often."
Source: "How we’re training a Figma-fluent editing model", Figma blog, Oct 29, 2026 (Josh Clemm) https://www.figma.com/blog/training-a-figma-fluent-editing-model/
Patrick, his desk, the product card, the two models as characters, the rubric and the counters are dramatizations; every quote and number on screen is from the source. Figma credits the gains to improvements across the agent, including the models, tools, and harness.
Would you trust a specialist model with your edits over GPT?
Transcript
Ask AI for one small fix. It breaks the layout. Undo. Undo. But it's not just you, Patrick. Figma's own article admits the top models couldn't even read Figma files at first. So Figma trained its own on over a quarter million example edits. Even an older Llama kept up with the top models back then. But copying has a ceiling. It repeats what it saw and stumbles on anything new. So now expert designers write its rewards. A button can look right and still be the wrong component. That's a miss. So they tested it. 445 real edits. Figma's model, 77.8. But GPT 6.1, 76.9. Under a point on Figma's own test. So Figma didn't pick one. New ideas, images, vectors. The best outside model, precise edits, their own. So since beta, tasks run 20% faster and people hit undo 60% less. The generalist dreams, the specialist fixes. Built for the moment when…
