Every language, with its score and its price
A language should not decide whether the tools work for you. We serve hundreds that no major model speaks, price each one by the model it actually needs, and publish the measurement behind every claim. No account, no API key, no permission to ask — pay the invoice and the work runs.
| Language | Model | Score | Price / 1,000 chars |
|---|---|---|---|
| Loading the live catalog… | |||
Scores are chrF measured on FLORES-200 against human reference translations. Prices move with Bitcoin, so the figures above are re-quoted hourly and the invoice is always the authoritative price. A language shown as “not benchmarked” has no independent test set — for us or for anyone.
The language picks the model
We route per language, from measured results, and take the cheapest engine within 2 chrF of the best. A difference too small to read is not worth paying a premium for.
We keep re-checking
Models improve and new ones ship. We re-run the comparison as independent results publish and move a language whenever something measurably better becomes reachable — at the lowest price that holds the quality.
Coverage is the point
Hundreds of these languages have no benchmark, no leaderboard and no score anywhere, because almost nobody serves them. We say so plainly instead of inventing a number.
Permissionless, in the literal sense
Most AI tools ask who you are before they ask what you need: an account, a card, a country they are willing to serve. That gate falls hardest on the people whose language is already the least served. Bitcoin removes it. There is nothing to sign up for and nobody to approve you — a payment either settles or it does not, and the same price is quoted to everyone for the same work.
That is why the long tail is not an afterthought here. Serving Wolof or Chhattisgarhi is not charity and it is not a rounding error in a roadmap — it is the part of the job most providers skip, and it is the reason this list is as long as it is.