OpenRouter ships seven-language localization in beta, with a daily self-correcting loop cutting glossary and authored-check failures from 147 to 4
Synopsis
OpenRouter's site localization is now in beta, adding Simplified Chinese, Spanish, Brazilian Portuguese, German, French, Japanese and Korean. The site holds 16,918 English strings that change daily, so a one-time translation would be stale within a week; a daily loop finds changed strings, translates them in batches, checks placeholders and glossary terms in code, and has the same model verify meaning, retrying with the reasons attached. On a 454-case pipeline comparison, the full pipeline cut glossary or authored-check failures from 147 to 4, while back-translation meaning retention stayed roughly level with the bare model.
Interpretation
OpenRouter's interface now offers seven languages beyond English, with localization in beta starting on launch day; users pick a language from the account menu or open a localized address such as openrouter.ai/ja or openrouter.ai/ko/models. The site has 16,918 English strings and some change every day, so a one-time translation would be out of date within a week. The daily loop handles only changed strings; a string enters the translation queue within a day of merging, and the page shows English until then.
Each translated entry stores a fingerprint of the English it came from, so when the English is edited the fingerprint stops matching and the entry returns to the queue, letting the daily run find every stale entry without retranslating the whole catalog. Translations are committed as JSON catalogs in Git next to the code and loaded locally at runtime, so no translation service sits in the request path. The pipeline has five stages: extract strings with their UI roles, translate in batches, check in code, verify meaning in a second call, and merge as a pull request by a person.
Code rejects replies that break an ICU placeholder, a plural branch, a rich-text tag or leading or trailing whitespace, or that use a glossary-forbidden term form; rejected batches are resent in smaller pieces with the failure named, up to 4 passes by default and a setting that goes to 20. Verification is a second call to the same model, which lists errors by severity, and any critical or major error sends the string back to the translator with every objection so far attached. The translator gets up to 3 retries after its first attempt, the last using more reasoning effort; counting check resends, a string can take about 20 model calls per language in the worst case, and if the verifier is down the pipeline writes nothing.
In the 454-case pipeline comparison, the bare model failed a glossary or authored check on 147 cases versus 4 for the full pipeline, and three judges from other model families counted 203 critical errors against 109. The bare model used a two-line prompt with no style notes, glossary or retries. In the same comparison, overall preference went 139 to 117 for the pipeline, with 61 ties and 136 identical pairs; back-translation meaning retention was 90.8% for the bare model and 89.3% for the pipeline.
Perspective
The demonstrated scope is OpenRouter's own frontend copy in seven languages, evaluated with a 254-case model comparison, a 454-case pipeline comparison and a 22-case defects suite; the cost figures cover translation and verification calls only, excluding engineering time to build the pipeline, AI reviewer and triage calls, and human merges.
Localization is still in beta and some wording will be off, which readers are asked to report through select-to-report; back-translation meaning retention was 89.3% for the pipeline against 90.8% for the bare model, which the team reads as a tie because back-translation cannot see terminology or register.
