The AI cost of JavaScript frameworks
Points and comments are a snapshot, not live.
A benchmark measures AI coding agent cost, speed, and quality across 12 JavaScript frameworks.
Nolan Lawson built a Pokémon app using agents (Opus 5.5) with 12 frameworks/meta-frameworks over 5 runs each. Hono+htmx was cheapest at $2.67 median cost, Astro scored 0 KiB JS and 100 Lighthouse performance. Angular was most expensive ($4.97), had the largest JS (93 KiB), and failed in 2 of 5 runs. SvelteKit agents wasted tokens reading migration guides. Next.js LCP varied from 1028-2187 ms due to image loading choices. The author notes the app is a simple content site, not a complex interactive app, and that Lighthouse biases against SPAs.
What commenters are saying
Top comments applaud Astro's results for content-driven sites, calling it pleasant and well-designed for zero-JS builds. A second camp argues the benchmark oversimplifies: frameworks should be judged on maintainability, team skills, and real-world complexity, not just agent build speed. Angular defenders blame the simple app type; detractors cite years of poor benchmarks. Several commenters advocate plain HTML or server-side rendering with minimal JS, saying frameworks add unnecessary boilerplate for agent-built apps. Some note SvelteKit's poor showing stems from agents reading v3 migration guides repeatedly.