TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A report published Oct. 28 compared a coding agent’s cost and app output across 12 JavaScript frameworks using repeated builds of a Pokémon data app. Hono with htmx had the lowest median API cost at $2.67 per run, while Angular had the highest at $4.97 among the frameworks tested; the author cautioned that results varied and covered only one kind of app.
Nolan Lawson published a benchmark on Oct. 28, 2026, comparing how much a coding agent cost to build the same web app using 12 JavaScript frameworks and related tools. Hono with htmx had the lowest median API cost, at $2.67 per run, while Angular had the highest reported median, at $4.97; Lawson said the results are specific to one app and varied across repeated attempts.
For the test, Lawson asked Claude Code running Opus 5.5 to build a small Pokémon information app connected to PokéAPI. The app included listings, details such as types and moves, and search and filtering. Frameworks included Next.js, Nuxt, SvelteKit, Astro, Angular, and several other options, alongside a vanilla JavaScript build and Hono with htmx.
Lawson ran the task five times per framework and reported median results. Angular was an exception: it failed to produce a measurable app in two attempts, leaving three iterations for some measures. The cost figures are the API prices shown by Claude Code, not money Lawson paid out of pocket; he said he used a monthly Max plan.
The reported median cost ranged from $2.67 for Hono with htmx to $4.97 for Angular. Astro came in at $2.76, Nuxt at $2.82, vanilla JavaScript at $2.98, and Next.js at $3.14. SvelteKit’s median was $4.28, and Gracile with Lit’s was $4.65. The report also examined development time, code size, and Lighthouse performance and accessibility scores, but these cost figures alone do not establish an overall winner.
What Framework Choice Costs Agents
The benchmark addresses a growing practical question: if coding agents generate much of an application, do framework choices affect the resources needed to produce it? In this single test, the spread between the lowest and highest median costs was $2.30 per run. That is a measurable difference for the tested task, but it is not evidence that the same gap will hold for larger projects or repeated production work.
For teams evaluating agent-assisted development, the report suggests that claims that a framework is inherently cheaper or easier for agents should be tested against their own workloads. API cost is only one part of that decision. The benchmark also considered spec conformance, development time, code volume, and Lighthouse scores; those measures can affect the usefulness and cost of an app beyond the agent’s bill. The supplied results do not establish that the least expensive build was best overall.
The findings also underline a challenge for comparisons: a difference between framework medians may be small relative to variation between individual agent runs. Lawson noted that, for example, React Router’s median was higher than Hono with htmx’s, but its cheapest run cost less than Hono with htmx’s most expensive run. Readers should treat the figures as results from a controlled attempt, not a price list for future projects.
Products Worth Considering
Apple Watch Series 12 GPS Smartwatch, 46 mm
LG UltraGear 32GX850A-B 32-inch 4K UHD Glossy OLED Gaming Monitor
Apple Watch Ultra 4 GPS + Cellular 49mm Smartwatch
As an affiliate, we earn on qualifying purchases.
How the Framework Test Worked
Lawson framed the work as a response to broad claims about which tools are “faster” or “better for agents,” saying such claims are often offered without benchmarks. He designed a content-focused Pokémon app because its requirements could be specified and checked. He acknowledged that it resembles a reference site more than a highly interactive product such as email or a design application.
To reduce differences in how the tools were presented, Lawson used each framework’s official MCP or skills where available and asked for the latest version. He did not dictate choices such as server-side versus client-side rendering, or ask the agent to optimize specifically for speed. The test prohibited Tailwind, Bootstrap, and other component frameworks, while using shared design tokens to limit visual variation.
Apps were evaluated against a conformance suite based on the accessibility tree rather than raw HTML, allowing different underlying structures, including web components using shadow DOM. Lawson also said the agents were not told that Lighthouse performance and accessibility tests would be used. These choices define what the benchmark measures—and what it does not.
Limits of the Cost Comparison
The findings do not show how framework costs compare across other app types, larger codebases, different specifications, or other coding agents and model versions. Lawson’s test used one agent setup and one relatively simple, content-heavy project; he did not prescribe an architecture or require performance optimization. Results may change with those choices.
Run-to-run variation also complicates rankings. Lawson said agents sometimes took unproductive paths and that Lighthouse measurements could differ between tests, including which page element counted as the largest contentful paint. The report used medians to reduce the effect of outliers, but the available results do not establish whether the observed differences would persist across broader or independently repeated studies. The excerpted findings also do not provide enough detail to determine which framework produced the best combined score across cost, conformance, accessibility, and performance.
Testing Results on More Apps
Lawson’s report presents the benchmark and its repeated-run results; the source material does not announce a follow-up schedule or a broader testing program. The immediate next step for readers applying the findings is to reproduce the task with their own requirements and compare multiple runs, rather than assume the reported medians predict their costs.
Further comparisons could test more kinds of applications, additional agents and model versions, and different rendering or optimization choices. Until such results are available, the figures are best read as a dated snapshot of one agent, one app, and one testing setup, not a general ranking of JavaScript frameworks for AI-assisted software development.
Key Questions
Which framework had the lowest reported AI build cost?
Hono with htmx had the lowest median reported API cost, at $2.67 per run, across five builds of the test app.
Which framework had the highest reported median cost?
Angular had the highest reported median, at $4.97. It produced a measurable app in only three of the attempted iterations, so its sample was smaller than the five-run samples used for most frameworks.
Did Lawson personally pay those API costs?
No. The report says the figures reflect API prices quoted by Claude Code and do not represent Lawson’s out-of-pocket spending. He said he used a monthly Max plan.
Does the benchmark prove one framework is best for AI development?
No. It tested one coding agent on one relatively content-focused app. The cost results do not establish which framework is best across other projects or when cost, performance, accessibility, and conformance are considered together.
Source: nolanlawson.com
