Test finds AI agents build the same web app cheaper in some frameworks
Developer Nolan Lawson had Claude Code with Opus 5.5 build the same small Pokémon web app five times in each of 12 JavaScript setups, then checked it against a test suite and Lighthouse. Median cost at API prices ranged from $2.67 for Hono with htmx, $2.76 for Astro and $2.82 for Nuxt, up to $4.28 for SvelteKit and $4.97 for Angular; Astro was fastest at 8.2 minutes and sent no JavaScript to the browser, against 142 KiB for Next.js. Performance and accessibility scores were near perfect for almost all of them, and agents spent extra tokens on frameworks too new for the model, such as reading the SvelteKit migration guide every run. The author stresses it is one small, content-style app and the results are noisy. On Hacker News (46 top-level comments when we read it) there is no consensus: several were pleased to see Astro, htmx or Nuxt do well or said plain HTML is enough, while others called the test too simple and said maintainability and team knowledge matter more than build cost. Sorted take: the method is more useful than the leaderboard.
Why it matters: If agents build your internal sites or tools, the framework you pick changes your token bill and build time. Before standardising, have your agent build one small real page in two candidate stacks and compare cost, time and code size.
Nolan Lawson ↗Discussion on Hacker News ↗AI-assisted, reviewed by Sorted AI
▲ 0
0 comments