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.

Simple, Composite, Complex tiers

Three complexity tiers beyond the single fixed component in the receipts — text-prompt only, run against both Claude and Qwen. These specs needed six blocks not yet exercised above (form, table, data-list, filter-bar, tabs, modal), and Claude's numbers here are "marginal" — total tokens minus the ~20,644-token shared harness overhead every Claude Code session pays, so they're comparable in kind to Qwen's raw completion totals. Run 2026-08-30 against rebar-ui v0.1.0.

Why this section only has n=5, not n=15

A single-run first pass here briefly looked like the two models disagreed on which tool wins. Repeating those runs showed that was just sampling noise, not a real disagreement — so we scaled up to n=5 per condition (n=3-4 on two cells that hit repeated technical failures, not silently dropped) before trusting the direction. Both models agree, all three tiers, once the sample is big enough — n=5 is still a smaller, less certain sample than the n=15 experiments elsewhere on this site.

Claude (marginal tokens, n=5)

Tier	Condition	Mean (marginal)	Difference
Simple	antd	8,104	—
	rebar-ui	7,999	−1.3%
Composite	antd	8,591	—
	rebar-ui	8,402	−2.2%
Complex	antd	8,988	—
	rebar-ui	8,712	−3.1%

Qwen (raw tokens, n=3-5)

Tier	Condition	Mean	n	Difference
Simple	antd	1,413	5	—
	rebar-ui	1,269	5	−10.2%
Composite	antd	2,138	4	—
	rebar-ui	1,881	3	−12.0%
Complex	antd	2,762	5	—
	rebar-ui	1,794	5	−35.0%

Both models agree once the sample is large enough

rebar-ui wins all three tiers on both models — a modest 1.3-3.1% on Claude, a much larger 10.2-35.0% on Qwen, the same bigger-gap-on-the-cheaper-model pattern found in the main n=15 experiments elsewhere on this site.

One representative build per condition (final run of each n=5 batch)

Claude · antd · Simple

1 of 12

Caveats: n=5 per condition (n=3-4 for two Qwen composite cells), one day (2026-08-29/30). All builds typechecked and were Playwright-verified (clean render, zero console errors beyond antd's already-noted deprecation warning). Claude was instructed to build one-shot (no file exploration, no self-verification) each run, confirmed genuinely one-shot afterward (exactly one tool call per run), to mirror what Qwen's raw completion call structurally can't do either way.