On November 12, 2026, OpenRouter announced on its official blog that its site now supports seven additional languages: Simplified Chinese, Spanish, Brazilian Portuguese, German, French, Japanese, and Korean — eight including English — currently in beta. You can switch in the account menu or open a localized address such as openrouter.ai/ja directly. The real news is not a few more languages, but the translation pipeline behind them that runs automatically every day.
A pipeline that retranslates itself every day
OpenRouter's site has 16,918 English strings, and they change every day, so a one-off translation would be out of date within a week. It built a self-correcting loop: every day it finds the strings that changed, translates them into seven languages, checks the format in code, and has a second model verify the meaning against the English; anything that fails is sent back with the reason attached to be translated again. Translation uses GPT-6.1 Sol, sent in batches of up to 40, with output locked to a strict JSON schema; code checks reject broken interpolation variables, plural branches, rich-text tags, or glossary-forbidden wording, with up to 4 resends; verification has the same model review the translation again in a different role, and serious errors are sent back too, with up to 3 retries and more reasoning effort on the last one. A string that cannot be translated stays in English — no unchecked translation goes live. Each day's translations land as a pull request, an AI reviewer reads the diff, and a person merges it. Readers can also highlight a translation on the page to report it, and the model triages it first: a wrong meaning is reverted to English the same day and retranslated, a wrong term goes into the glossary, and a tone problem goes into that language's style notes.
Why the cheaper model was replaced
The choice was made on evaluation data. The cheaper model tried first, GPT-6 Luna, passed 91.7% of 254 cases on the first try, against 99.2% for GPT-6.1 Sol; Luna was weakest in Korean and Japanese, and some of its errors could not be fixed by any number of retries. The expensive tier cost $1.94 for those 254 cases, the cheap one $0.57 — about 3.4 times more — and OpenRouter still picked the expensive one: fixing one bad translation afterward costs more in human time than the price difference. The pipeline's effect can also be measured: across 454 cases, the bare model translating directly produced 147 glossary or hand-written check violations, the full pipeline only 4; with three outside model families acting as judges, critical errors fell from 203 to 109. Retranslating the whole site into seven languages costs about $460 in total, about $0.0005 per word; at the starting rates Smartling and Gengo publish, the same job would cost $7,200 to $86,000; only changed strings are translated each day, and a day with 677 changed strings costs about $18. The system is a necessity for OpenRouter itself: it only just completed its acquisition by Stripe on November 10, and the unified gateway Stripe bought for $7.5 billion has to serve users worldwide — an English-only interface would not hold up. The other reminder is that OpenRouter previously measured the same task burning 42 times more tokens over MCP than over a CLI: toolchain costs are never just the model price. The eight-language beta is live now; Chinese users can switch straight to the Simplified Chinese version, report mistranslations through the page prompts, and fixed rules go into the glossary so the same mistake does not recur.
