Price‑against‑score chart for LLMs tops Hacker News, readers dispute method
The site plots models on a value frontier from Artificial Analysis data, and the top comment calls the pricing estimate naive.
A single-page site plotting language models against their API prices reached 112 points and 67 comments on the Hacker News front page on Thursday. It draws a value frontier: the models for which no cheaper alternative scores as well on intelligence, coding or maths.
The data comes from Artificial Analysis's free data API, fetched daily by a GitHub Actions cron, according to the page. Prices use a 3:1 input-to-output blend per million tokens, and exclude cached-input discounts, batch pricing and fast modes. The maker gives the GitHub handle terryds and publishes the code and data.
The most-upvoted comment disputes the approach. Readers are better off going straight to Artificial Analysis, it says, calling the site a feature-poor and misleading repackaging. The pricing estimate is naive, that commenter argues, because some models may need two to three times the tokens to reach the same level of intelligence. Artificial Analysis's own cost-per-task figure is the fairer estimate, they wrote.
The page carries a caveat of its own. Scores move, it says, because Artificial Analysis rebases its index between versions, so comparisons should only cite the current page rather than an older snapshot. A minimum score threshold, 30 by default, filters out weak budget models that would otherwise bend the frontier line.
Much of the thread went elsewhere, to running models locally. One commenter said Qwen3.8 27B runs well overnight on a 24-to-64GB Mac, at Q4 on smaller machines, and ranks above GPT-5.3 Codex and Claude Opus 4.6 for their purposes. None of those comparisons has been independently checked.