Claude Sonnet 5.5 was the main session. Claude Opus was used as research and planning agents (xhigh and high effort). Claude Sonnet 5.5 agents implemented the plans.
It wrote the code, researched sources and planned the design.
The idea, the design direction, the data choices and the text.
The numbers in the language list are estimated, live; modelled, not counted. Each language starts from its total number of speakers (first plus second language) in Ethnologue 2026, as listed on Wikipedia and read on 17 November 2026. Ethnologue gives these totals in millions. The figures are dated 21 February 2026, Ethnologue's usual yearly release date; this date is an assumption, as the exact date is not stated.
Gains (births among speakers and new learners) and losses (deaths and people who stop using the language) arrive as two independent random (Poisson) processes. The loss rate is the crude death rate of the main countries where the language is spoken (UN World Population Prospects 2024, values for 2023), weighted by a rough share of speakers. The gain rate is the loss rate plus the yearly growth. The growth is the trend between the Ethnologue editions of 2022 and 2026 where that trend is smooth; where an edition jumps because of a new way of counting, it is the UN projected population growth for 2026 of the same countries.
The page shows the expected value plus a random fluctuation that is computed from the clock, so every visitor sees the same number at the same moment. Nothing is downloaded for this.
Pointing at a language in the list opens a pop-up with every country that has at least 2% of its speakers. The per-country numbers are modelled estimates in the same way, each country with its own starting figure, growth and loss rate; sources in provenance/country_speakers.md and provenance/country_speakers10.md.