A benchmark finds some web frameworks cost more for AI agents

A web developer made one Pokemon app with AI agents in many JavaScript frameworks. The benchmark checks the claim that some frameworks are good for agents. Each framework got 5 tests of Opus 5.5 set to high, and the median was the result. Angular did not make an app that works in 2 tests. The developer names limits of the test.

How it works

Each agent got the same spec for an app with Pokemon data. It used the MCP (a tool for agents) or skill files (written documents) of its framework. A test suite tests the output. Lighthouse, a tool that scores page speed and access for all users, tests the result.

The numbers

• The React Router median cost about $1 more than Hono with htmx.
• A low React Router test cost $2.82. A high Hono test cost $3.19.
• SvelteKit agents read a migration document in each test.
• Gracile agents cloned its source code. Angular cost follows the quantity of code.

What is not known

The app is one content site with a small quantity of interaction. Results are not the same each time. Agents used the tools of the framework in only some tests: the Next.js skill files in 1 of 5. The text has no results for quality or maintenance.

Insight: Agent claims for a framework need a measured cost

Agents did not use the Nuxt MCP in no test. They found TanStack Start skill files in 2 of 5 tests. SvelteKit agents read a migration document in each test, maybe because the model did not know the new version. Makers of frameworks seem to compete to be easy for agents, but a tool that agents do not use does not help. Before you change a frontend for agents, test the cost.