Published 2026-09-14 · updated 2026-09-25 — part of an ongoing build diary of dated, real measurements, not a one-time proof of the platform's superiority.

A third, genuinely different model: Kimi-K3

Qwen3.7 is one model family. To test whether the pattern above holds for a real architectural difference, not just another Qwen size, the same text-prompt spec was run against Kimi-K3 (Moonshot AI, via the same DashScope-compatible endpoint) — a reasoning model, unrelated to Qwen. Same rigor as every condition above: n=15 per condition, every run written to disk and type-checked, a Playwright-verified subset for clean render and correct DOM order, full screenshot galleries. Run 2026-08-30 against rebar-ui v0.1.0.

Condition	Mean	Median	Min	Max	Std. dev.	Success
antd	5,967	5,931	3,520	9,456	1,821 (30.5%)	15/15
rebar-ui	1,925	1,889	1,578	2,869	278 (14.5%)	15/15

Wall-clock	Mean	Median	Min	Max	Std. dev.
antd	160.0s	155.6s	70.0s	261.5s	56.2s (35.1%)
rebar-ui	37.5s	32.5s	25.4s	72.8s	12.4s (33.0%)

antd

1 of 15

rebar-ui

1 of 15

The gap is bigger here than on Qwen — and much bigger than on Claude

rebar-ui costs 67.7% less on Kimi-K3 (vs. 63.8% on Qwen3.7-max, 3.1-3.3% on Claude) and finishes 76.6% faster. A third model, a different vendor, the same direction and an even larger gap — one more real data point that removing layout/composition decisions helps more the further a model is from frontier-level open-ended authoring, not proof of a smooth scaling law across three data points.

antd's cost on Kimi-K3 is genuinely unpredictable — rebar-ui's isn't

Look past the mean: antd's worst Kimi-K3 run cost 2.7x its best one (3,520 to 9,456 tokens) for the exact same spec — build the identical UI twice and you might pay nearly three times as much the second time, for no reason you'd be able to predict or control. rebar-ui's own spread over the same 15 runs was much narrower (1,578 to 2,869, a 1.8x worst-to-best ratio) — still real variance, just far less of it. This is the same "pixel-consistency" story told in tokens instead of pixels: a wider, less predictable range isn't just a cost problem, it's a planning problem — a hand-authored antd build's price is a lot harder to budget for than a rebar-ui one, on this model.

Caveats: n=15 per condition, 2026-08-30, same spec and blocks as the Claude and Qwen text-prompt experiments. Kimi-K3 was chosen after checking a couple of other third-party models reachable on the same platform — one produced no usable output at a normal token budget, another worked but cost noticeably more per call — so this is the one that held up. All 30 final runs typechecked and a Playwright-verified subset confirmed clean render and correct DOM order. antd's output shows the same deprecated Alert message prop found on every other antd condition on this page — harmless to rendering, a sign of antd v6 API drift the model's training hasn't caught up to.