The AI cost of JavaScript frameworks

The article describes a benchmark study testing how efficiently AI coding agents build a Pokémon app using the PokéAPI across JavaScript frameworks including Next.js, Nuxt, and SvelteKit, evaluating development time, tokens spent, lines of code, and Lighthouse performance and accessibility scores, with spec conformance testing to prevent agent cheating. The author argues that web development decisions remain "vibes-based" rather than evidence-driven, even in the agentic era where claims that certain frameworks are "better for agents" are made without benchmarks. Each framework used its official MCP or skills to level the playing field, with the spec given to Opus 5.5. The author, who references Alex Russell's "frontend's lost decade" critique, notes caveats that the app is a single holotype—a relatively content-based site with little interactivity, closer to Wikipedia than GMail—and that architecture choices like SSR vs CSR were not specified. The piece matters because it challenges framework marketing claims with empirical data as AI-assisted development reshapes tooling choices. What to watch: full benchmark results across frameworks and whether the community adopts evidence-based framework selection for agentic workflows.