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.

Does iteration change the winner?

Everything measured elsewhere on this site is one build. The real question a single build can't answer: once realistic follow-up work (add a field, reorder content, add a note, change an option) is layered on round after round, does antd's head start survive, or does rebar-ui's pay-once-at-migration cost structure eventually overtake it? Rather than stop at an arbitrary round count, each tier was iterated until it actually crossed over — a measured answer, not an extrapolation. Run 2026-08-30 against rebar-ui v0.1.0.

Condition A — antd direct. Build the UI against Ant Design components, then apply one follow-up prompt per round, repeated until crossover.

Condition B — rebar-ui, then migrate once. Build the same UI headless with rebar-ui, apply the identical follow-up prompts, then reimplement it directly in antd once at the end (the migration step — see below for why this isn't a codemod run here).

Simple — A settings form — a handful of fields, one submit action.
Composite — A list view with filters, a modal, and inline validation.
Complex — A multi-step wizard: table + form + confirmation dialog + tabs.

Rigid by design: rebar-ui is meant to refuse style/color requests

rebar-ui is deliberately rigid — closer to rebar and formwork in real construction than a flexible styling toolkit: it builds the load-bearing structure (logic, accessibility, content) correctly, then refuses to let that structure be visually fine-tuned before the "cladding" goes on, once, at migration. So when the first follow-up prompt here tried a pure restyle ("make this button softer, add padding"), two of three rebar-ui agents refused it outright — working exactly as intended: the placement schema has no per-instance color/style override, and inventing one would mean fabricating an API that doesn't exist. The third was explicitly told it could edit shared files to satisfy the request — overriding that rigidity on purpose, to see what happens. What happened is exactly the failure mode the rigidity exists to prevent: it made the restyle work, but by editing shared packages/core CSS and BlockRenderer.tsx, and in the process made a second, undisclosed, unrelated change to a shared button style. That edit was reverted immediately; nothing from it is in the numbers below, and the library's own docs now state the refusal as an explicit rule rather than leaving it to an individual agent's judgment. Every round below was redesigned around content/structure changes instead (add a field, reorder, add a note, change an option, cycling through those four types) — requests both architectures can actually express.

Cumulative marginal cost per round (Claude, harness-subtracted)

"rebar-ui + migration" adds the one-time migration cost to round 0, so the two lines are comparable at every round, not just at the end. Green marker = the round where rebar-ui + migration first becomes cheaper than antd.

Complex — crosses over at round 13

Composite — crosses over at round 14

Simple — crosses over at round 17

All three tiers cross over — between round 13 and round 17

Complex crosses first (round 13: antd 138,439 vs rebar-ui+migration 137,756), Composite next (round 14: 142,981 vs 142,584), Simple last (round 17: 167,673 vs 166,797). The more complex the spec, the sooner migration pays for itself — consistent with rebar-ui's per-round saving being roughly proportional to how much content a round touches, while the one-time migration cost only grows sub-linearly with complexity. Not a smooth countdown: round 15 on the Simple tier briefly widened the gap instead of closing it (a single round's real cost swung the other way that time) before resuming its close — a visible reminder that any one round is still an n=1 measurement, even though the multi-round trend across three independently-run tiers is consistent.

Final state (post-crossover), one representative build per condition

antd · Simple (round 17)

1 of 6

Caveats: n=1 per round per condition (97 dispatches total across all three tiers) — a single round's number can swing (see round 15's temporary reversal above), so the exact crossover round for any tier could land a little earlier or later on a repeat. Migration here is a full LLM rewrite to equivalent antd code, not a partial codemod pass, so it's a real, fully-counted cost, not an underestimate. Every final state typechecked and was Playwright-verified.

A structural head start worth naming: models haven't been trained on rebar-ui

Every number measured on this site compares a library the model has seen constantly in training (antd) against one it has never seen at all (rebar-ui, and the @rebar-ui/placement schema specifically). That's not a flaw in the comparison — reaching for a small, complete unit needs far less training familiarity to use well than assembling one from a large component API does, which is a real advantage of the tier system independent of training exposure. But it does mean today's numbers likely understate rebar-ui's ceiling, not overstate it — the image-prompt experiment shows a real, measurable "unfamiliar schema" tax (extra lookup calls before the prompt was refined to spell the schema out directly). If a future model were actually trained on examples of this schema, that residual tax would plausibly shrink further, on top of the advantage already measured here — a directional expectation, not something this page has tested.