Same app, 12 frameworks: Claude Code costs nearly twice as much in Angular, and ignores the MCPs
Framework choice now has a price tag in agent dollars. Nolan Lawson had Claude Code build the same app in 12 JavaScript frameworks, five runs each, and checked every result against a spec conformance suite so the agent couldn't cheat its way to done. The post went up October 28 and ran on the Hacker News front page this weekend with 267 points.
Median cost per build, at API prices: Hono with htmx was cheapest at $2.67, then Astro $2.76, Nuxt $2.82, plain vanilla JS $2.98 and Next.js $3.14. At the expensive end, SvelteKit cost $4.28, Gracile with Lit $4.65 and Angular $4.97, with Angular runs ranging from $3.19 to $6.82. Part of SvelteKit's cost came from agents reading the v3 migration guide.
The finding that matters more for agent builders: each framework's official MCP server or skill was installed, and the agent mostly ignored it. The Nuxt MCP was never called. Next.js agents read the skill file in one run out of five and AGENTS.md in two. TanStack Start's skill got found in two of five.
Vendors are shipping MCP servers and skill files on the assumption that agents will use them. This is a small, careful test showing that installed isn't the same as used. If you maintain a framework or an API, the question is no longer whether you have an MCP but whether an agent will pick it up without being told. Lower-ceremony frameworks are already cheaper to build with.
Post: https://nolanlawson.com/2026/10/28/the-ai-cost-of-javascript-frameworks/ Repo: https://github.com/nolanlawson/agent-framework-benchmark
← Back to all articles
Median cost per build, at API prices: Hono with htmx was cheapest at $2.67, then Astro $2.76, Nuxt $2.82, plain vanilla JS $2.98 and Next.js $3.14. At the expensive end, SvelteKit cost $4.28, Gracile with Lit $4.65 and Angular $4.97, with Angular runs ranging from $3.19 to $6.82. Part of SvelteKit's cost came from agents reading the v3 migration guide.
The finding that matters more for agent builders: each framework's official MCP server or skill was installed, and the agent mostly ignored it. The Nuxt MCP was never called. Next.js agents read the skill file in one run out of five and AGENTS.md in two. TanStack Start's skill got found in two of five.
Vendors are shipping MCP servers and skill files on the assumption that agents will use them. This is a small, careful test showing that installed isn't the same as used. If you maintain a framework or an API, the question is no longer whether you have an MCP but whether an agent will pick it up without being told. Lower-ceremony frameworks are already cheaper to build with.
Post: https://nolanlawson.com/2026/10/28/the-ai-cost-of-javascript-frameworks/ Repo: https://github.com/nolanlawson/agent-framework-benchmark
Comments