Which LLM writes correct UI?

Each model generates the same 5 interfaces, sampled 3 times each, rendered in headless Chromium, and graded with deterministic checks only: data fidelity, axe-core accessibility, layout, console errors. No LLM judges another LLM. Rerun on every model release.

generated 2026-07-14 · uivet 5c077ed · N=3 runs/scenario · temperature 0

Ranking

#	Model	Provider	Passed	Fidelity	A11y crit+serious	Latency
1	Gemini 3.5 Flashgemini-3.5-flash	google	5/5	100%	3	37.0s
2	Gemini 3.1 Flash-Litegemini-3.1-flash-lite	google	5/5	100%	6	3.0s
3	Gemini 2.5 Flashgemini-2.5-flash	google	5/5	100%	17	23.7s
4	GPT-5 minigpt-5-mini	openai	4/5	100%	6	35.4s
5	GPT-5gpt-5	openai	5/5	100%	1	53.2s
6	GPT-4.1 nanogpt-4.1-nano	openai	4/5	100%	10	8.5s

Per scenario

Model	flight-results	expense-form	metrics-dashboard	pricing-table	settings-form
Gemini 3.5 Flash	PASS 100%	PASS 100%	PASS 100%	PASS 100%	PASS 100%
Gemini 3.1 Flash-Lite	PASS 100%	PASS 100%	PASS 100%	PASS 100%	PASS 100%
Gemini 2.5 Flash	PASS 100%	PASS 100%	PASS 100%	PASS 100%	PASS 100%
GPT-5 mini	FAIL 100%	PASS 100%	PASS 100%	PASS 100%	PASS 100%
GPT-5	PASS 100%	PASS 100%	PASS 100%	PASS 100%	PASS 100%
GPT-4.1 nano	PASS 100%	PASS 100%	PASS 100%	PASS 100%	FAIL 100%

Method

A scenario passes iff mean fidelity is 100% (every value in the scenario data appears in the rendered page text across all runs) and there are zero axe-core critical violations.
Reported, not gated: serious axe violations, layout flags, console errors, mean generation latency.
Ranking: scenarios passed, then fidelity, then a11y critical+serious, then latency.
Raw per-run data: results.json. Reproduce: bun leaderboard/run.ts.