Methods
Where every fact comes from, how each one is checked, what AI does in the process and how popularity is estimated.
What we record
The record holds 545 changes from 35 sources about 13 makers. Every change comes from a page that an AI research agent opened and quoted, and goes through the checks in the next section. These are the pages we read:
- The makers' own changelogs, release notes, news posts, pricing pages and deprecation pages.
- Official status pages, for outages and errors.
- Changelogs of coding agents and open source projects, such as the ones published on GitHub.
- Independent test results, such as Terminal-Bench, Arena, Epoch AI and Vals AI.
- Independent quality trackers, such as MarginLab and Livenerf.
- Hacker News discussions, for problems that users report.
We record an outage only from an official status page, and only when it is rated major or critical or lasts 30 minutes or longer. Every quality report carries one of three labels: "Confirmed by the lab", "Measured" (an independent test that repeats with a stated method) or "Reported, unverified" (user reports, shown with the size of the discussion).
Each fact on a profile, such as a price, a context size or a benchmark score, links to its own source page, and the profile shows the date it was researched. The short headline on each profile, such as "Flagship Engineer at Anthropic", is written from those sourced facts.
A change reaches the site once a maker's page or one of the independent sources above records it, so a change announced only on social media appears when it is written up there. You can follow every recorded change by RSS, and each changelog page has its own feed at its address plus /rss.xml.
How every change is checked
- Every four hours, AI research agents read each source for anything new since it was last read. For every change they draft the date, what changed, who it affects, what to do about it, and a short quote copied word for word from the page.
- Code checks the format of each change and removes duplicates.
- Code downloads every source page again and confirms that the quote, and every number in the change, appears on it. A change whose quote cannot be found is corrected from the page or dropped.
- A separate AI review, which did not write the change, scores it against the source page: that the change says what the page says, that it names the right model and version, and that the source is reliable for this kind of claim. Claude Sonnet scores every change, Claude Opus scores it again when the result is close or the change reports a security or quality problem, and OpenAI's GPT-6.1 Sol also checks any serious accusation. Only changes that score well are published.
Because every step is code or a model with fixed instructions, each change gets the same checks at any hour, and a new change can appear within hours of its source. The quote check confirms the words are on the page, and the score confirms that the change says what the page says. If you find a mistake, report it through the contact form. We fix it and log the fix on the corrections page.
Comparison pages are built the same way. Agents research each pair, a script confirms every quote on its source page, a separate AI review scores the result, and the summary is written only from those sourced facts.
What AI does here
- AI research agents find changes and draft their text, research profile facts and draft comparisons. They run on Anthropic's Claude today. The same sources and checks decide what we publish about every model, Claude included.
- Code checks every quote and number against its page and computes popularity. A separate AI review scores what the agents drafted before it is published.
- Because the agents run on Claude, Claude also drafted the text on the Claude profiles. Those pages go through the same checks as every other page.
- The character illustrations were made with OpenAI's GPT Image model from prompts we wrote. Each one is original art, not a company's logo or mascot.
How popularity is estimated
Search data checked Oct 24, 2026
The "Most popular" order uses how often people in the US search for each name in a month, as estimated by DataForSEO from clickstream data. Follower counts will replace it once there are enough followers.
- Months with less than 5 percent of a name's best month are left out, so a new release is not held back by its quiet first weeks. We then take the middle value of the last three months that remain, so one unusual month does not move the order.
- Searches for a whole line, such as "claude opus", are shared equally by its current versions.
- A version that has been replaced counts a quarter and always ranks below the version that replaced it. A version with a published shutdown date counts half, and a retired one counts zero.
- A brand name that also means other things, such as "kimi", only limits what one version can score.