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.
Claude Sonnet 5 (n=15 per condition)
rebar-ui composes the target through a compact schema plus the Packer, a deterministic renderer built from rebar-ui's own components; antd is hand-authored directly. Same prompt, same model, isolated scaffolds, only the build step measured. What this isolates is assembling from primitives versus reaching for a complete unit, not the schema wrapper itself — see the Coherence field evidence for the caveat. Run 2026-08-29 against rebar-ui v0.1.0 — see the Kanban benchmark for the same underlying question re-measured against v0.12.1.
Text prompt
| Condition | Mean | Median | Min | Max | Std. dev. |
|---|---|---|---|---|---|
| antd | 31,231 | 31,131 | 30,891 | 31,921 | 294 (0.9%) |
| rebar-ui | 30,211 | 30,212 | 30,149 | 30,253 | 28.6 (0.095%) |
| Wall-clock | Mean | Median | Min | Max | Std. dev. |
|---|---|---|---|---|---|
| antd (n=14 of 15) | 24.5s | 23.7s | 21.6s | 30.1s | 2.4s (9.7%) |
| rebar-ui | 14.7s | 14.6s | 14.2s | 15.8s | 0.47s (3.2%) |
antd
rebar-ui
Image prompt
Same two conditions, same target — but each run was handed the reference screenshot below instead of a written spec, and had to read the layout off it directly.
| Condition | Mean | Median | Min | Max | Std. dev. |
|---|---|---|---|---|---|
| antd | 31,765 | 31,659 | 31,021 | 34,257 | 776 (2.4%) |
| rebar-ui | 30,787 | 30,769 | 30,684 | 31,073 | 118 (0.38%) |
| Wall-clock | Mean | Median | Min | Max | Std. dev. |
|---|---|---|---|---|---|
| antd | 27.2s | 25.4s | 20.8s | 54.4s | 8.0s (29.6%) |
| rebar-ui | 16.7s | 16.4s | 15.0s | 20.2s | 1.4s (8.1%) |
antd
rebar-ui
Caveats: n=15, one model (claude-sonnet-5), one day (2026-08-29), one component spec. All runs type-checked; a Playwright-verified subset (clean render, zero console errors, correct DOM order) confirmed both extremes and the median for each condition. antd's image-prompt output uses a deprecated Alert message prop (antd v6 renamed it to title) — harmless to rendering, a sign the model's antd knowledge lags the v6 release this benchmark targets.