AI Translation

Send text and a target language, get a translation back — structured JSON, not prose you have to strip a preamble off.

POST /api/v1/translate
X-Api-Key: ksty_live_...
Content-Type: application/json

{"text": "Good morning", "target_language": "Japanese"}

{
  "translated_text": "おはようございます",
  "romanization": "Ohayō gozaimasu",
  "detected_source_language": "English",
  "target_language": "Japanese",
  "usage": { "input_tokens": 38, "output_tokens": 24 }
}

Source language is optional

Leave source_language out and the model detects it from the text. Send it when you already know it — a short or ambiguous phrase is easier to translate correctly with the source language pinned down than guessed at. Either way, detected_source_language in the response tells you what the text was actually treated as, so a caller who guessed wrong can see that rather than getting a translation from the wrong starting point silently.

Romanization

A translation into a non-Latin script comes back with romanization alongside it — the translation sounded out in Latin letters, using the standard scheme for the language (Hepburn for Japanese, Pinyin with tone marks for Mandarin, Revised Romanization for Korean). It is the difference between a string you can display and a string you can say.

romanization is always present and is an empty string when there is nothing to sound out:

• the translation is already in Latin script (French, Spanish, Vietnamese), or
• text is longer than 500 characters.

That second limit is a billing decision rather than a technical one. A romanization roughly doubles the output tokens, and output is the expensive half of this API — worth it for a phrase someone is about to say, wasteful on a document nobody will read aloud. Send long text in short pieces if you do want it romanized throughout.

What "faithful" means here

The model is instructed to translate what is there — including phrasing you might not expect a translation service to touch — rather than to edit, soften, or summarise. Formatting (line breaks, lists, punctuation) is preserved as closely as the target language allows. This is a translator, not a rewriter: if you want the meaning adapted rather than the words translated, that is a different task from what this endpoint does.

Pricing — why this is metered per token

Like text-ai and ocr, this calls a paid language model, so it is priced the same way:

Unlike text-ai, where a short summary is cheap regardless of document length, translation's output is close to its input in size — sometimes longer, since some language pairs expand text by 15-30%, and roughly double on a short call that also returns a romanization. The gateway's pre-charge accounts for that with a generous per-call ceiling; X-Credits-Charged is what the call actually cost once the difference is refunded.

Live keys only, and no per-API free allowance

translate has no test sandbox and no monthly free grant, for the same reason as text-ai and ocr: every call spends real money on a metered model upstream. A ksty_test_ key gets 403 sandbox_disabled. Calls draw on your normal account balance, which includes the free credits released with your first API key.

Limits

• 20,000 characters of text per call, about 8 pages. Split longer documents and translate the parts — unlike a summary, a split translation loses no accuracy from being chunked.
• One target language per call.

When to use it

• Translating support messages, product listings, or UI strings on demand rather than maintaining a static translation table.
• Localising user-generated content where a fixed phrase list will not cover what people actually write.
• Pairing with text-ai's language parameter when you want a translated summary instead of a translated original.