When AI writes the code, your framework choice has a price tag

Nolan Lawson had Claude Opus 5.5 build the same app in 12 JavaScript frameworks, five times each. The cheapest framework cost a median $2.67 per build and the most expensive $4.97, at API prices. For a founder paying for an MVP, that gap is small. The bigger differences in his results are how much JavaScript each framework sends to the user, from 0 KiB to 142 KiB, and how reliably the agent finished the job. Pick the stack for the kind of app you're building, then ask your developer how they check what the agent writes.

I build MVPs and web apps for founders, and this site runs on Next.js. Lawson's benchmark drew hundreds of comments on Hacker News this week, and it's the first public test I've seen that puts numbers on a question clients now ask me: does the framework matter if AI writes most of the code?

How did Nolan Lawson test the AI cost of frameworks?

Lawson gave Claude Opus 5.5, at high effort, the same task in each framework: a small but non-trivial Pokémon app built on the PokéAPI, with search and filtering. The agent used each framework's official MCP server or skills where one existed. He ran each framework five times and reported the median, then checked every result with a conformance suite and Lighthouse on a throttled mobile profile.

The 12 frameworks were Angular, Astro, Gracile with Lit, Hono with htmx, Next.js, Nuxt, Qwik, React Router, SolidStart, SvelteKit, TanStack Start and plain Vite.

Which framework was cheapest and fastest for the agent?

Two results surprised me. Every framework scored 100 for accessibility, so the agent got the basics right everywhere. And plain Vite with no framework did worse than several frameworks, which suggests a good framework still saves the agent work. Lawson notes the costs are API-equivalent prices, since he ran the tests on a subscription plan.

Does the framework change what your MVP costs?

Barely, at build time. A $2 difference per run is noise next to the hours a developer spends planning, reviewing and testing. If an agent builds a feature ten times while you iterate, you're still talking about tens of dollars. The test I use for whether AI spend pays off is in why AI pays off for some companies and not others.

The costs that do add up come later:

• What your users download. Astro shipped no JavaScript for this app. Next.js shipped 142 KiB and had the slowest largest contentful paint. On a content-heavy site, that is page speed you pay for in search rankings and conversions.
• How reliably the agent finishes. Angular had two failed runs, and the agent built server-rendered apps that rejected outside requests. Every failed run is time someone spends finding out what went wrong.
• How much code there is to maintain. Hono with htmx did roughly the same job as Angular in about 70% of the code. Less code means less for the next developer, or the next agent, to read.

Which framework should you choose for your MVP?

Choose by the kind of app, then by evidence. Lawson's test used a content-heavy app with little interactivity, and he points out that this favours lightweight, multi-page approaches. Your app may not look like his.

My rule of thumb from the projects I take on:

• Mostly content, a few forms: a marketing site, a directory or a docs site. A light framework such as Astro ships less JavaScript and loads faster.
• An app people log in to: dashboards, account areas, AI features with live data. React-based frameworks such as Next.js earn their weight here, because the interactivity is the product.
• An app you started in Lovable or Bolt: keep the stack it was built on unless it's failing. A rewrite usually costs more than fixing what's there.

The one thing I'd never do is pick a framework because it's what the developer always uses. Lawson asks the same question: "Are you really picking your tech stack based on evidence? Or are you just going with what you already know?"

What should you ask a developer about their stack?

Ask these four before you sign off on an MVP:

1. Why does this framework suit my kind of app?
2. How much JavaScript does the first page send, and how fast does it load on a phone?
3. How do you check and test the code an AI agent writes before it ships?
4. If you stopped working on this tomorrow, could another developer or agent pick it up?

A good answer to question 3 matters more than the framework. Clear prompts help too. PrePrompt rewrites a prompt with the context and constraints an agent needs, which cuts down the reruns that cost tokens.

I build MVPs and web apps from $2,400 to $15,000, and I finish and harden apps started in Lovable or Bolt from $2,000, on my published pricing. Tell me what you're building and I'll tell you which stack I'd use and why.