Best AI models for translation

What to weigh for translation

Translation reads and writes the same amount, so both price columns count, and it is one of the tasks where general models have overtaken dedicated systems. The models above $5 blended are excluded here: translation quality has plateaued close to the frontier for the major languages, and the frontier price buys little.

The intelligence index tracks translation quality well enough to rank by, with two cautions: for less common languages the gap between models widens, and tone - formal or informal address, product terms - is an instruction-following question the better-scoring models handle more reliably.

The router treats translation as an action it can detect in the request, so a rule can send it to a dedicated skill or a specific model and leave the rest of the catalogue out of it; glossaries and style live in the skill, not in every prompt. The language itself is a signal too - a request written in French or Tamil can be matched to the models that handle it well - and the outcome loop catches the rest: a translation your reviewers mark wrong moves the next one in that language to the stronger model while the common languages stay on the cheap one.