Angular Costs an AI Agent Twice as Much as Hono+htmx: $4.97 vs. $2.67 per App

When a team chooses a framework, it usually thinks about developer habits, the ecosystem, and hiring. Now there is another factor: how much it will cost to have an AI agent write code in place of a human. Developer Nolan Lawson decided to measure this in dollars and minutes.

How the experiment was set up

The task was the same for everyone: a Pokémon catalog with search, filters, and real data from PokéAPI. It was carried out by Claude Code running the Opus 5.5 model with a high reasoning level. Lawson tested 12 variants, including plain JavaScript with no UI framework. He ran each variant five times and took the median, so that random luck or failure would not skew the picture. The agent received the same task description and functional tests used to verify the result.

The method is simple, which is why it deserves more trust than individual impressions like "Claude writes React well." Here there is a common bar and repeatability.

What came out

Angular had the worst results on both metrics at once: 17.2 minutes and $4.97 per app. Astro finished fastest, needing just 8.2 minutes. The cheapest was the Hono and htmx combination: $2.67 and 8.6 minutes. Angular took roughly twice as long and cost twice as much as the leaders.

Plain JavaScript turned out decent but not the best: $2.98 and 10.1 minutes. It beat neither Astro nor Hono with htmx. The thesis "fewer abstractions means it's easier for the agent" was not confirmed here. Apparently, what helps the agent is not the absence of a framework, but clear conventions and compact code that doesn't take long to comprehend.

In tokens, the spread is as follows: Nuxt spent 3.85 million, Angular 9.76 million, that is, 2.5 times more. Most of the spending came from repeatedly reading cached context. The longer the "write, run, fix" loop, the more often the agent rereads the accumulated history, and the bill grows almost on its own.

Quality barely suffered

The unexpected result is that the final apps differed little. Median Lighthouse performance was 97–100 points, and accessibility was 100 everywhere. So the expensive framework did not deliver a better product. It only required more steps and tokens for the same result.

What to do with this

You shouldn't choose a stack based on these numbers alone. This is one app, one model, and one scenario. A catalog with search and filters fits a server-first approach easily, while a complex enterprise frontend with forms, state, and dozens of screens may paint a different picture. Besides, token prices and model behavior change quickly.

Still, there are practical takeaways:

• If you run agents on a large volume of tasks such as prototypes, internal tools, or landing pages with dynamic elements, a 2x difference turns into real money. It's worth running your own benchmark on your typical tasks.
• For quick prototypes built by an agent, it makes sense to consider Astro or Hono with htmx. They delivered the best time and price at the same quality.
• If Angular is already adopted in your team, that's no reason to abandon it. Just budget for the agent working longer and costing more, and keep an eye on context size.
• The metric should be counted as a whole: tokens, minutes, and the share of repeated context reads. Optimizing only the price per request is not enough.

The main takeaway: a framework now affects the cost of automated development as well. That cost is worth measuring the same way as build speed or bundle size.