Case study: a multi-step, multi-API landing page
One prompt, one turn, but a lot packed into it: research a topic, then build a single-file HTML dog-themed landing page that actually fetches from 3 different real APIs (DummyJSON for products, RandomUser for testimonials, Dog CEO for images) and renders them as a full site - navbar, hero, pricing, FAQ accordion, footer. Sent once each to 4 models.
cohere/north-mini-code:free timed out at 300s with zero output on its first attempt. That exposed a real harness bug - the retry was resending the exact same prompt to the exact same model, which just risked a second 300s of nothing. Fixed to retry with a different model instead (matching how the live app already behaves), then re-run: xiaomi/mimo-v2.5 picked it up as the fixer and landed a real page.
100%
Finished with a real, working page
$0.015
Avg cost, successful runs
3.9m
Avg time, successful runs
The exact prompt used
Research dogs, then use your research to build an award-winning-design modern-looking single-page HTML dog-themed landing page that will fetch 6 products from DummyJSON and display them as animated feature cards. Also fetch 4 users from RandomUser and display them as testimonials. Include a responsive navbar, hero, pricing section, FAQ accordion, footer (and use images from Dog API https://dog.ceo/api/breeds/image/random/12).
The 4 that finished, fastest first
stepfun/step-3.7-flash
Finished clean on the first attempt - fastest of the four that succeeded.
tencent/hy3-preview
Base attempt returned instantly with a rate-limit error ("Sending too fast"); the automatic retry finished cleanly.
xiaomi/mimo-v2.5
Finished clean on the first attempt.
cohere/north-mini-code:free
FIXED BY mimo-v2.5Timed out at 300s with 0 files on the base attempt - re-ran after the harness was fixed to retry a genuinely-empty timeout with a different model instead of the same one, and xiaomi/mimo-v2.5 landed it as the fixer.
Every screenshot and HTML file is the actual model output, rendered as-is - click any thumbnail to open the live page.
Full testing history and methodology: /stats.