Field SOP
Field SOP

OpenRouter: 16,918 Strings, About $460, 7 Languages

Breaks down the OpenRouter site localization pipeline from the official announcement: 16,918 English strings, a full retranslation into 7 languages for about $460 including verification, and three distinct language counts across the body, the FAQ, and the index. The verification model and the two feedback channels stay separate.

Published November 14, 20269 min read
<!-- openrouter-i18n-pipeline-sop | sop | OpenRouter: 16,918 Strings, About $460, 7 Languages -->

As of November 13, 2026, the OpenRouter announcement page (byline Robert Sun, dated November 12, 2026) lays out the site's localization work as a set of figures that can be checked line by line. The site has 16,918 English strings, in the official wording OpenRouter's site has 16,918 English strings. A full retranslation of 16,918 strings into 7 languages costs about $460, translation and verification included, and the post adds our pipeline costs about 6% as much, a comparison the official page draws against the Smartling machine-translation price. This piece reproduces the pipeline's wording verbatim from that official post and lists only dimensions a reader can verify, without adding UI paths of its own. Every figure below is tied to a phrase in the official post, so a reader can check it against the source rather than take this piece's word for it. Two angles are worth separating up front. This is not the same as the earlier OpenRouter single-task token-cost write-up, which measured how many tokens one GitHub task consumes under MCP versus CLI and is about a single task's cost; this piece is about site-wide retranslation and the engineering pipeline's cost basis. It is also different from the AI translation tools comparison, which weighs six tools against each other; this piece does not compare tools, it reconstructs one in-house pipeline.

7 Languages, 8 languages, and where each count comes from

The language count is not one number in this official material; three phrasings coexist and each is kept as written. The body of the post repeatedly says 7 languages, and that is the set of translation targets: English strings get rendered into 7 languages, and the cost math of 16,918 strings times 7 languages follows that basis. The FAQ lists the interface languages as English, Simplified Chinese, Spanish, Brazilian Portuguese, German, French, Japanese, and Korean, which is 8 interface languages once English is counted. The blog index summary reads, verbatim, OpenRouter now speaks 8 languages., and the announcement URL slug is openrouter-in-8-languages; that 8 languages is the wording of the index and the address. Note that the full body of the post does not contain the string 8 languages, so the index phrasing should not be turned into a blanket claim about the site's total language count, and the translation target of 7 should not be mistaken for the interface-language total. Seven is the target set, the FAQ is 8 interface languages including English, and the index and URL say 8 languages: three counts, kept side by side rather than flattened into one. Localization is in beta, and that is as far as the official status wording goes; this piece does not infer when beta ends.

The five-step pipeline, as verifiable dimensions

The official overview of the whole chain is this: new and changed strings go out in batches, then checks them in code, then a second model verify the meaning against the English, and finally An AI reviewer reads the diff and posts findings, and a person merges the pull request. When it gets to the Verify step and the FAQ, the same post writes the same model for that checking role. Both phrasings are kept as written, in separate sentences, not merged and not renamed. Below, the pipeline is expanded in five steps, namely batch translation, code checks, verification and retries, AI review with human merge, and reader-report capture, and each step lists only what a reader can check. Verifiable here means that, holding the official post, a reader can tell what kind of constraint a step describes, rather than having buttons, menus, or console paths invented for an interface the full page never lays out in that detail.

Batch translation. What can be checked: only new and changed strings are sent, not the whole site re-pushed each round. The official wording is We send new and changed strings to GPT-6.1 Sol in batches of up to 40, with the model ID openai/gpt-6.1-sol and at most 40 strings per batch. The scale of 16,918 strings across 7 languages is the base for the cost math below. Localization is in beta, so this is a starting state rather than a long-running finished process; keeping the status separate from the scale is what stops a one-off full retranslation bill from being read as a recurring daily cost.

Code checks. The official post summarizes this as checks them in code, keeping deterministic checks on the code side: format, placeholders, and strings inside code are the kind of checks with a clear right or wrong that are not handed to a model to improvise. The verifiable dimension is telling apart what belongs to hard constraints decidable in code from what needs a model to judge meaning. A model is good at meaning; whether placeholders stay paired, whether variables get mistranslated, whether format markers get eaten are better caught by code first. Putting hard constraints in code up front is also what leaves the later model calls for the places that genuinely need semantic judgment, and it is the premise for pricing verification separately from translation. The full page gives this step a single sentence, so no buttons, menus, or console paths are added here.

Verification and retries. In the Verify step, the official wording is the same model, checking the translated meaning against the English. The verifiable dimensions are a set of caps: up to 4 passes by default (with the ceiling set to 20), up to 3 retries, and about 20 model calls per string per language. Verification on its own is priced at verification about $228, or $0.002 per string per language; translation and verification together give the earlier about $460, translation and verification included. These figures are kept verbatim and not recomputed: 4 passes is a default cap, 20 is a ceiling, 3 is a retry count, and about 20 model calls per string per language is the official scale description, with none of them restated as a derived percentage or a converted total.

AI review and human merge. The official line An AI reviewer reads the diff and posts findings, and a person merges the pull request splits the last gate in two: an AI reviewer reads the diff and posts findings, while a person still merges the pull request. The verifiable dimension is that human-machine boundary, how far automation goes and which step stays with a person. This is distinct from the reader-report line below: issues caught by the daily review become test cases and feed a defects suite (22 cases) as regression cases, and that is the pipeline's own daily automated review, not outside feedback, with a different entry point and a different destination.

Reader-report capture. This is a separate feedback source running alongside the daily review, with the entry point the official page calls select-to-report in the Language dialog. When a reader reports a translation problem, the official handling is we turn it into a glossary rule, a style note, or an override: turn it into a terminology rule, a style note, or a direct override of that string, so the same class of problem does not depend on luck next time. The verifiable dimension is whether such feedback lands in reusable rules rather than just patching one string in front of you. The two sources stay separate: one comes from readers and goes to a glossary rule, a style note, or an override; the other comes from the daily AI review and goes to test cases and a defects suite (22 cases). They are not the same mechanism and are not merged into one sentence.

Two feedback sources, not one mechanism

It is worth pulling the fifth step out again, because this is where the two lines most easily get folded into a single sentence. Reader reports are outside input: a reader selects a stretch of translation in the interface and reports a problem, and the official page turns it into a glossary rule, a style note, or an override, a long-lived constraint at the terminology and style level, with the entry point select-to-report in the Language dialog. The daily AI review is inside input: an AI reviewer reads the diff each day, and issues it catches become test cases collected into a defects suite (22 cases), an accumulation of regression cases. One faces readers, the other faces the pipeline itself, and the terms for where each lands differ; the official page writes them separately, and so does this piece, without merging them or calling them one mechanism. The usual cost of conflating the two is treating a reader opinion that belongs in a terminology rule as a regression case, or treating a defect the daily automation caught as a one-off override. Keeping each to its official wording is what lines up with the separate roles of a defects suite (22 cases) and a glossary rule.

The cost ledger, in the official numbers

Laid out, the scale and cost are a string of directly checkable figures. Baseline batch: 16,274 strings fully retranslated at about $430, yielding 113,918 translations; word count from 136,429 words to 955,003 words, which the official page puts at about $0.0005 per word. The 16,918 strings basis reads: A full retranslation of 16,918 strings into 7 languages costs about $460, translation and verification included, with verification alone about $228, or $0.002 per string per language. Incremental cases: 677 changed strings cost roughly $18; 97 such strings (0.09%) are the small handful needing extra handling. For comparison, the official page lists prices as of November 10, 2026, and our pipeline costs about 6% as much is said against the Smartling machine-translation tier:

TierUnit priceFull-volume equivalent
Smartling MT$0.0075/wordabout $7,200
Smartling AI$0.06$57,300
Gengo Standard$0.09$86,000
Smartling AI Human$0.12$114,600
Gengo Advanced$0.18$171,900
Smartling Human$0.20$191,000

These are the comparison numbers in the official announcement, copied item by item without summing them up or restating their basis; about $460 and the earlier about $430 are different batches and should not be blended into one. About $460 is the figure most readers remember, but the verification share of about $228 is what makes the translation and verification included wording concrete. On model choice, the official page also tested GPT-6 Luna (91.7% vs 99.2%), with Claude Opus 5.5, Gemini 4 Argon, DeepSeek V4 Pro as judges; those are self-test figures for verification quality, a different dimension from the cost figures above, and this piece only reports the numbers the official page gives rather than drawing conclusions for it. Separating cost, verification quality, and language count is what keeps a single about 6% from turning the whole announcement into a plain price-cut claim.

FAQ

Q1: Does the official page support 7 languages or 8 languages? A1: Three counts coexist, each kept as written. The body's 7 languages is the translation target set. The FAQ lists 8 interface languages including English: English, Simplified Chinese, Spanish, Brazilian Portuguese, German, French, Japanese, and Korean. The index summary reads OpenRouter now speaks 8 languages., and the URL slug is openrouter-in-8-languages. The full body does not contain the string 8 languages, so the index phrasing is not a blanket statement about the site's language count, and the target set of 7 is not the interface total.

Q2: Which model verifies the translations? A2: In the Verify step and this FAQ's wording, the official page writes the same model, matching translated meaning against the English; the pipeline overview uses a different phrasing, and both are kept separately rather than merged into one sentence or renamed. The main translation model is GPT-6.1 Sol (openai/gpt-6.1-sol), sent in batches of up to 40.

Q3: What happens when a reader reports a translation problem? A3: The official wording is we turn it into a glossary rule, a style note, or an override: a terminology rule, a style note, or a direct override of that string; the entry point is select-to-report in the Language dialog. This is outside feedback from readers, and the verifiable dimension is whether that class of problem gets folded into reusable rules rather than only patching the one string in front of you.

Q4: And where do issues caught by the daily AI review go? A4: The official wording is become test cases, feeding a defects suite (22 cases) as regression cases. This is inside input from the pipeline's daily automated review, a different source and a different destination from the reader reports in the previous answer, not the same mechanism and not merged into one sentence.

Q5: What does about $460 include, and what is about 6% as much compared with? A5: Verbatim, A full retranslation of 16,918 strings into 7 languages costs about $460, translation and verification included, so translation and verification are both inside it, with verification alone about $228. The phrase our pipeline costs about 6% as much is against the Smartling machine-translation price as of November 10, 2026, a tier whose full-volume equivalent is about $7,200; the earlier 16,274 strings batch was about $430 and is a different batch from the 16,918 strings used above, so the two should not be mixed.

Related reading

This article is AI-assisted and human-edited. Last updated: 2026-11-14

FAQ

Does the official page support 7 languages or 8 languages?
Three counts coexist, each kept as written. The body's 7 languages is the translation target set. The FAQ lists 8 interface languages including English: English, Simplified Chinese, Spanish, Brazilian Portuguese, German, French, Japanese, and Korean. The index summary reads OpenRouter now speaks 8 languages., and the URL slug is openrouter-in-8-languages. The full body does not contain the string 8 languages, so the index phrasing is not a blanket statement about the site's language count, and the target set of 7 is not the interface total.
Which model verifies the translations?
In the Verify step and this FAQ's wording, the official page writes the same model, matching translated meaning against the English; the pipeline overview uses a different phrasing, and both are kept separately rather than merged into one sentence or renamed. The main translation model is GPT-6.1 Sol (openai/gpt-6.1-sol), sent in batches of up to 40.
What happens when a reader reports a translation problem?
The official wording is we turn it into a glossary rule, a style note, or an override: a terminology rule, a style note, or a direct override of that string; the entry point is select-to-report in the Language dialog. This is outside feedback from readers, and the verifiable dimension is whether that class of problem gets folded into reusable rules rather than only patching the one string in front of you.
And where do issues caught by the daily AI review go?
The official wording is become test cases, feeding a defects suite (22 cases) as regression cases. This is inside input from the pipeline's daily automated review, a different source and a different destination from the reader reports in the previous answer, not the same mechanism and not merged into one sentence.
What does about $460 include, and what is about 6% as much compared with?
Verbatim, A full retranslation of 16,918 strings into 7 languages costs about $460, translation and verification included, so translation and verification are both inside it, with verification alone about $228. The phrase our pipeline costs about 6% as much is against the Smartling machine-translation price as of November 10, 2026, a tier whose full-volume equivalent is about $7,200; the earlier 16,274 strings batch was about $430 and is a different batch from the 16,918 strings used above, so the two should not be mixed.

Related

Field SOP

OpenRouter: MCP Costs 17x More, Tool Results Are the Culprit

A measurement published by OpenRouter on November 11, 2026 wired one cli/cli repository question three ways: on Claude Sonnet 5.5, the default MCP server used 42 times the CLI input tokens and cost 17 times as much, $127 against $7.50 per 1,000 tasks. The official TL;DR points at tool results rather than tool definitions as what fills the context, with 20,400, 42,182, and an 8,336 to 714 trim all checkable row by row. This is a single-task measurement and a self-check list, not an integration guide or a price comparison. Tool Search varies by model: it lowers the full-task total on Sonnet while making Haiku and Gemini more expensive, and it does not shrink tool results. Run mcp_cost.py on your own tasks before deciding, as the official post advises.

Nov 13, 20269 min read
Field SOP

Self-host the open-source Dots in three commands

A self-hosting SOP for the open-source Dots: install uv (Windows PowerShell and Linux routes), then one uvx command brings it up (--openrouter-key required); open 127.0.0.1:8765 for the chat-left, live-browser-right workbench. Swap models with one --model flag across anything on OpenRouter; four advanced parameters explained in depth - --seed (one identity per seed, reproducible), --profile-dir (logins and cookies persist across runs), --proxy (timezone and language follow the exit) and --help (the full option index); invisible_playwright_mcp hands the same browser to Claude Code, Codex and Gemini CLI as an MCP server; the README's own sample task (Milan-to-Lisbon daily fare check, no-ticket means say so) demonstrates the no-fabricated-numbers prompt discipline. A closing compliance section: respect target-site terms and local law, automate only your own business and accounts you are authorized to use, no fraud/scalping/bulk sign-ups. All commands and parameters come verbatim from the README.

Oct 2, 20268 min read