Every model takes a limited number of tokens at once (its context window), and Thai costs more tokens than English. Paste a document or enter a character count to see which models take all of it.
The most tokens a model takes in one exchange — instructions, attached documents, chat history and the answer it writes. That is why an answer reserve is subtracted.
Thai tokens per character differ a lot between models, so a bigger context does not always hold more Thai.
Fitting is not the same as reading well. “Lost in the Middle” (Liu et al., TACL 2023) found that the models of the time used information in the middle of long inputs worse than at the start or end. For very long documents, put the question and the key parts first or last.
How it is calculated
Thai tokens per character were measured on 109 real Thai articles on this site, which mix in English terms like typical tech news. Pure Thai may take slightly more.
GPT, Qwen and Typhoon use their real tokenizers; the other families use ratios measured through OpenRouter against GPT (see the token counter page).
Context sizes come from OpenRouter on every site rebuild; latest 2026-10-21. Some providers serve a smaller context than the model supports.
With a pasted document: GPT is counted exactly; the rest are estimated from your text’s share of Thai letters.
