A consequence worth adding for anyone building on top of this: tokenization is not uniform across languages, and that turns a curiosity into a budget line. Text in Turkish, Finnish or any agglutinative language fragments into far more tokens than the equivalent English, because the vocabulary was fitted mostly to English text - so the same document costs noticeably more, fills the context faster, and gets truncated sooner. It also degrades retrieval quality in a way that is easy to blame on the wrong component, since heavier fragmentation means embeddings for those languages are built from less coherent pieces. Same underlying fact as the strawberry bug, just with a bill attached.