Impeccabenchv1

November 20268 models · 13 briefs

Can a model turn a product brief into a website that a designer would ship? Each model gets the same brief, the same tools and the same image generator, and builds the page from scratch. We measure the page it hands back, with and without the Impeccable design skill.

PREVIEWScores come from an automated judge that has not been calibrated against human ratings yet. Treat small gaps as noise.

Leaderboard

Models without any design skill, ranked by the automated judge. The second score shows the same model with Impeccable installed.

#Model Human Elo
blind pairs		 Judge
no skill		 LandingApps Judge with
Impeccable		Lift FinishedCost
per page		Minutes
Judge scores are on a 1–5 scale: the weighted mean of six design-critique categories (typography, color, layout, imagery, polish, distinctiveness), with hard caps from deterministic checks. Finished is the share of runs that handed back a working page. A run that produced nothing, timed out or stalled counts against the model and is left out of its score. Cost and minutes are per page without the skill, at list prices on the day of the run.

How it works

The briefs

Thirteen product briefs: seven landing pages and six application screens. Each brief describes a product and its audience, not a design. The model starts in an empty project and has to build a single working page.

    The run

    Every model runs in its own vendor's agent harness at default settings, with a browser to look at its own page and the same image generator for photos and illustrations. The brief is the only input: if the model asks questions, a simulated user tells it to use its own judgment. Each model runs every brief two or three times, and every run counts.

    The scores

    An automated judge (Gemini 3.8 Flash) scores a screenshot and the page's code against six design-critique categories, and deterministic checks cap scores for measurable faults such as overflowing text or unreadable contrast. The judge has not been calibrated against human taste yet, and it is itself a Google model, so read Gemini's rows with that in mind. Human Elo comes from blind side-by-side ratings by a designer and replaces the judge as the headline once there are enough of them.

    With and without the skill

    The headline ranks the models on their own. The second score installs Impeccable, a design skill for coding agents, in the same harness and runs the same briefs. The difference shows what a skill adds on top of each model.