Technology news and discussion on TalkFuse.
Pokémon app test finds server-centric JavaScript frameworks cheaper for AI agents
Nolan Lawson benchmarked 12 JavaScript frameworks by asking Claude Code to build the same Pokémon app with data from PokéAPI. The results suggest AI coding agents can work more efficiently with server-centric frameworks than with client-heavy single-page apps, especially when tools favour established conventions, stable APIs and concise code.
Lawson ran five sandbox trials per framework, then checked the apps with automated accessibility tests and Google Lighthouse. Hono with htmx had the lowest median build cost at $2.67, followed by Astro at $2.76; Astro also had the fastest median build-and-test time, at 8.2 minutes. Nuxt was the least costly option among the SPA-like approaches at $2.82 and used the fewest tokens, 3.85 million.
In short
- Hono with htmx had the lowest median cost at $2.67.
- Astro built and tested the app fastest, in 8.2 minutes.
- Server-centric tools generally suited the AI agent better.
Steelman: both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
The benchmark suggests server-centric frameworks can let AI agents produce working apps with less code, fewer tokens and lower costs, particularly when conventions and stable APIs make tasks easier to follow. For teams prioritising economical, fast development, the Hono with htmx and Astro results offer a practical reason to favour that approach.
The case against
The results come from one developer’s specific Pokémon app task, one AI coding tool and five trials per framework, so they may not predict performance on other projects or with other agents. Teams may reasonably prioritise client-heavy frameworks for their interaction model, ecosystem or existing expertise, and Nuxt’s results show that an SPA-like option can also be cost-effective.
Explained simply
JavaScript frameworks provide common ways to build websites and apps. Some send finished pages from a server, while others send code to the browser and update the page there; these are often called single-page apps. Htmx is a tool for adding interactive features to web pages, and PokéAPI is an online source of Pokémon data.
Nolan Lawson compared 12 frameworks by having Claude Code, an AI coding agent, build the same Pokémon app with data from PokéAPI. He repeated each build in a sandbox and used automated checks to assess the results. The comparison matters because the choice of framework may affect how much time and computing resources an AI agent needs to build an app.
Read the full article at the source →
Originally published by The Register as “Pokémon bakeoff reveals best JavaScript frameworks for AI”.