The AI cost of JavaScript frameworks

Claims that one framework is better for agents rarely come with evidence, so Nolan Lawson built some. Claude Opus 5.5 (high effort, in a sandbox) built the same small Pokémon browser five times in each of 12 setups, from Angular and Next.js to Hono+htmx and plain Vite, checked against an accessibility-tree conformance suite. Median cost ran from $2.67 (Hono+htmx) and $2.76 (Astro) to $4.97 (Angular), with Gracile+Lit and SvelteKit ($4.28) close behind; cost, time and tokens moved together. Server-rendered and multi-page builds did better on Lighthouse, and Hono+htmx matched Angular's features in about 70% of the code. One detail hits close to home: this site runs on SvelteKit 3, and the agent read the SvelteKit v3 migration guide in every run, presumably because the release is newer than the model's training data. Lawson also suspects today's framework MCP servers and skills do more harm than good. He is candid about the limits: one small content-heavy app, one model, Chromium only, a largely vibe-coded harness. The code is on GitHub and he invites rebuttals.