theknowngood.com tracks benchmarks, pricing and ratings for ~900 AI models

What launched

A new site, theknowngood.com, is live and collects published evaluation data for AI models in one place: benchmark results, pricing, performance figures and arena ratings. According to a dev.to post by its creator, the site covers around 900 models across roughly 2,300 pages, and its CSV and JSON exports are free, ungated and published under a CC BY 4.0 licence. There is no account system, no user tracking and no API.

The author describes themselves as a systems engineer rather than a developer, and says the decisions worth discussing in the project are architectural ones.

A static site rendered from a database

PostgreSQL is the source of truth for the data, but nothing queries it while someone is browsing. Python and Jinja2 templates render the entire site into flat HTML files on a schedule, so a visitor opening a model page is simply reading a file from disk.

The creator gives two reasons for this design. The first is reproducibility, because a static build can be regenerated and compared between versions. The second is cost, because serving flat files stays cheap under load in a way that generating pages per request does not.

No ports open to the internet

The hosting arrangement is deliberately unusual. The nginx web server binds only to the machine's loopback interface, and the site reaches visitors through an outbound tunnel, meaning nothing listens on the public internet at all.

In the dev.to post, the author argues that this eliminates a whole category of security problems outright rather than trying to defend against them, a trade they prefer as the only person on call for the service. The site runs on hardware in their own home, and they invite visitors to report slowness directly rather than leaving it to surface in logs weeks later. An optional newsletter on the homepage requires no account.

Charts that disclose what they measured

One data-presentation rule the author highlights: every chart carries its own denominator. A top-N ranking shows how many models were actually evaluated for that chart, not the total size of the catalogue. The author calls this the most important correction made during the entire project, despite sounding minor at first.

The post also points to a write-up of everything that went wrong while building the site, including a long period in which most of the built-in checks were passing while effectively measuring nothing.

Why it matters

Evaluation data for AI models is scattered across papers, leaderboards and vendor pricing pages, which makes like-for-like comparison slow and error-prone. A single, openly licensed dataset covering roughly 900 models, downloadable as CSV or JSON and reusable under CC BY 4.0, lowers the barrier for researchers, tool builders and anyone weighing models against one another.

The architecture is also a useful reference point for solo maintainers and small teams. Static generation keeps hosting costs and page-load times low, while the loopback-plus-tunnel setup removes public exposure rather than patching it. The denominator discipline speaks to a familiar criticism of leaderboards, where rankings can appear to describe a much broader population than was ever actually evaluated.

The limitations are just as clear. This is a one-person, home-hosted project with scheduled rebuilds, so freshness and uptime rest on a single operator. Whether the data stays current will determine whether theknowngood.com becomes a durable reference or an interesting snapshot.