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.

Qwen3.7 (n=15 per condition)

Same spec, same blocks, same rigor — run against Qwen (qwen3.7-max for text, qwen3.7-plus for image) via the DashScope API, to test whether a cheaper model benefits even more from having layout/composition decisions removed from it. Run 2026-08-30 against rebar-ui v0.1.0.

Not directly comparable to Claude's numbers

Qwen's numbers come from a single raw completion call — no agentic tool use, no file access, none of the ~22,000-token harness overhead a Claude Code session pays on every run. That's why Qwen's totals (roughly 1,000-4,000 tokens) look so much smaller than Claude's (30,000+) — it's a thinner slice of work, not a more efficient model. The fair comparison is within Qwen's own results: antd vs. rebar-ui, same model, same harness.

Text prompt

Condition	Mean	Median	Min	Max	Std. dev.	Success
antd	3,388	2,989	1,597	6,446	1,376 (40.6%)	15/15
rebar-ui	1,228	1,065	994	2,028	278 (22.6%)	15/15

Wall-clock	Mean	Median	Min	Max	Std. dev.
antd	46.6s	40.3s	19.7s	88.3s	20.8s (44.6%)
rebar-ui	10.3s	9.4s	6.0s	21.7s	3.9s (37.8%)

antd

1 of 15

rebar-ui

1 of 15

Image prompt

Condition	Mean	Median	Min	Max	Std. dev.	Success
antd	3,948	3,667	3,243	5,882	726 (18.4%)	15/15
rebar-ui	2,956	2,965	2,839	3,129	90.3 (3.1%)	15/15

Wall-clock	Mean	Median	Min	Max	Std. dev.
antd	33.2s	29.2s	20.5s	64.6s	11.9s (35.8%)
rebar-ui	13.8s	13.9s	11.2s	17.6s	1.7s (12.1%)

antd

1 of 15

rebar-ui

1 of 15

The relative gap is bigger on the cheaper model

On Claude, rebar-ui won by 3.1-3.3% on tokens. On Qwen, the same comparison shows a 25-64% advantage — a much larger relative win, in the direction the hypothesis predicted: removing layout/composition decisions helps more when the model doing the composing is weaker at open-ended authoring to begin with. Two models, two conditions each, is a real data point in that direction, not proof the effect scales smoothly with model capability.

Caveats: n=15 per condition, one day (2026-08-29/30), same spec and blocks as the Claude experiments. All 60 runs were written to disk, type-checked, and a Playwright-verified subset confirmed clean render and correct DOM order for both extremes and the median per condition. All 60 were also screenshotted for the galleries above, using the same pixel-alignment method as the Claude galleries. antd's output shows the same deprecated Alert message prop found on Claude's antd runs. A prompt written for a general-purpose coding agent doesn't automatically work for a single completion call with no file access — one condition needed its prompt rewritten to be fully self-contained before all 15 runs passed cleanly; full account in ref/QWEN_BENCHMARK_PROTOCOL.md for anyone reproducing this.