Twelve frameworks, one app: agent build cost ran from $2.67 to $4.97
Serguey Asael Shinder2 min readNolan Lawson, who has long argued for measuring frontend claims instead of repeating them, published a benchmark on 28 October that asks a question now common in framework debates: is a given framework "better for agents"? He wrote a spec for a small Pokémon browser built on the PokéAPI, had an agent (Opus 5.5 at high effort, with each framework's official MCP server or skills installed) build it with twelve frameworks, five runs each, and checked every result against a conformance suite based on the accessibility tree.
Build cost, taken as the median of five runs 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, $4.65 for Gracile with Lit and $4.97 for Angular, which failed to produce a measurable app twice. Lawson attributes part of the SvelteKit cost to agents reading the v3 migration guide on every run, and the Gracile cost to agents cloning the repository or reading compiled sources because they did not know the framework.
Lighthouse barely separated the results: performance scored 97 to 100 and accessibility 100 everywhere. The shipped JavaScript did: Astro sent a median of 0 KiB on the home page, the vanilla build 7 KiB, Hono with htmx 18 KiB, SvelteKit 36 KiB, Nuxt 84 KiB, React Router 115 KiB, TanStack Start 118 KiB and Next.js 142 KiB.
What it means
The useful part is the method more than the ranking. Lawson notes the measurements are noisy (the cheapest React Router run beat the most expensive Hono run), the app is one small, content-heavy case, and Lighthouse favours multi-page apps. Within those limits, two findings are sturdy. First, how well a model already knows a framework shows up directly in tokens, because unfamiliar or very new versions send the agent off to read documentation or source. Second, the agent did not erase the familiar size differences between frameworks: the JavaScript cost lands on users whichever way the code was written. A team weighing a rewrite "for agents" now has a template for running the comparison on its own app.