The Starborn design benchmark
Faster.
Better.
Cheaper.
We gave Gemini Flash reusable expert themes and three focused visual-review passes, then compared it with Opus designing from scratch. The goal: close the quality gap without giving up Flash’s speed or price. Here is all of the evidence.
The evidence
Judge the pages yourself.
Scores are out of 70 across hierarchy, typography, composition, polish, responsive behavior, objective WCAG contrast, and fit to the brief. Every Flash–Opus pair was blinded and judged three times. Click any image to inspect it full-size.
Measured, not estimated
The cost of each finished page.
| Brief | Checklist-guided Flash | Opus from scratch | Flash advantage |
|---|
One Starborn credit is one cent. Totals use each model’s reported tokens and include Starborn’s standard multiplier. Different output lengths and turn counts make the observed ratio vary by task.
How we tested
A useful benchmark should be easy to disagree with.
So we show the work. Flash selected one of seven reusable design systems; the server copied that complete CSS into the project, so Flash did not spend output tokens reproducing it. Three inexpensive screenshot-driven passes then focused separately on scope and composition, typography and art direction, and contrast and mobile polish. A quality ratchet rejected any pass that broke behavior or worsened measured contrast. Opus built a fresh version from the same brief. Every result was rendered at 1440×1000 and 393×852, shuffled, and judged three times by Gemini Pro.
The judge is directional, not divine. Scores vary. The pages, measurements, and comments are here so you can form your own view.
- 01GenerateFlash and Opus build from the same brief
- 02RenderDesktop and mobile screenshots
- 03BlindModel identities hidden from the judge
- 04MeasureQuality, contrast, seconds, credits
Build at Flash speed
Start cheap. Ship fast.
Keep making Flash better.
Starborn gives fast models expert design systems without paying them to regenerate thousands of CSS tokens. We publish the gap instead of hiding it—and use the feedback to close it.
Build your first page free