An open notebook on AI-made apps · Experiment 01

Same idea.
Different models.

Which model would you choose to build your next idea? Explore the same briefs, compare the results, and decide for yourself.

Loading the experiment…

Explore the results ↓ How we ran it ↗

Results, not a victory lap. Loading model coverage… Cohorts ran on different dates and app revisions. Timing includes tool work and interruptions—not just model speed. There is no overall winner.

Read before ranking

A comparison you can inspect.

Same prompt, paired results

The Flash batch covers 24 ideas and three prompts each. Pro and Opus add a small pilot: four ideas, prompt 1 only. All use the same app-building platform, on different dates and revisions. This compares whole agent workflows, not raw model intelligence.

Timing with context

The displayed duration is recorded wall-clock time. Browser pauses, infrastructure failures and recovery affected this batch. Do not use these numbers as a model latency ranking. Controlled reruns are needed.

Beauty is not functionality

These are screenshots, not a certification that every feature works. Capture blocks write requests, which can affect rendering. Functional, accessibility and safety checks must be scored separately. Private project data and access tokens are not published.

Your taste, your shortlist

Rate visual quality from 1–5. Your ratings stay in this browser and are not public votes or editorial scores. A useful scorecard separates visual quality, working features and performance by category—not just one overall number.

Download the public experiment data ↓

First pilot observations · Prompt 1 only

Different strengths, in these examples.

Research: an editorial hierarchy

Paper Circle's Opus result scored 8.7/10 for its readable journal layout and clear paper-opening action. Inspect the screenshot—not a claim that its citations or AI answers are correct.

Compare Paper Circle →

Workshops: character with a clear action

Clay Club's Flash 3.8 result scored 9.3/10: distinctive craft imagery alongside visible workshop details and a request button. Booking behavior has not been tested.

Compare Clay Club →

Recipes: two strong approaches

Flash 3.6 and Opus both scored 8.7/10 on Recipe Lab. One exposes useful task controls; the other emphasizes photography and browsing. Try your own visual ranking.

Compare Recipe Lab →

Private workspaces: don't score the gate

Both premium Proof Room captures show sign-in screens. We retained them without application scores. Authentication is not a design failure; a signed-in review is still needed.

Compare Proof Room →

Fourteen visual scores, two abstentions, one AI reviewer. These are exploratory observations from a tiny sample—not recommendations that a model is best for an entire category.

The right model depends on the job.

A research workspace and a restaurant landing page ask different things of a model. Browse by category, inspect the prompts, and look for the differences that matter to you.

Try your own idea →

Same idea.Different models.