Configure Immersive Translate's custom Chat Completions API

Set Immersive Translate's custom API endpoint and key

In Immersive Translate, open Settings → Translation Services → OpenAI and switch the service from Pro to Custom API Key. Paste your Router One key into APIKEY. Click Expand for more custom settings and enter the Custom API interface address. The field's own help text says it accepts a base URL or a full endpoint URL, and the full Chat Completions URL is the one value every version handles the same way. Then open Set up more models, add the catalog IDs you want (next section) and select one as the Model. Click Verify service, translate a short page, and match the new entries in Dashboard → Logs by time and model before you translate a whole document:

Custom API interface address（自定义 API 接口地址）

https://api.router.one/v1/chat/completions

# Immersive Translate → Settings → Translation Services → OpenAI
# 沉浸式翻译 → 设置 → 翻译服务 → OpenAI
Service（服务）:  Custom API Key（自定义 API Key）, not Pro
APIKEY:  sk-your-router-one-key
Custom API interface address（自定义 API 接口地址）, works in every version:
  https://api.router.one/v1/chat/completions
Set up more models（设置更多模型）:  -all,+anthropic/claude-haiku-4.5,+deepseek-v4-flash
Model（模型）:  pick one of the IDs you added, then click Verify service（点此测试服务）

Which address to enter: full URL, /v1 or bare host

Immersive Translate rewrites what you type into a request URL, and the rule changed between versions. The Chrome Web Store serves 1.33.3 (Edge Add-ons lists 1.33.1), and 1.33.3 completes almost anything; Firefox Add-ons still serves 1.30.2, which is stricter about a bare host. Enter the full URL and the version does not matter. An address that ends in /responses keeps that path, and Router One serves /v1/responses only for GPT-family models, DeepSeek IDs and Grok chat models, so stay on /chat/completions.

Add catalog model names with Set up more models

The OpenAI service's built-in model list holds OpenAI's own names without a vendor prefix, such as gpt-5.5 and gpt-5.4-mini, so add the exact catalog IDs yourself. Set up more models (设置更多模型) takes a comma-separated list: a name adds it (a leading + does the same), -name hides one of the built-in names, -all hides all of them, and name=Label gives a name already in the list a friendlier label. For example, -all,+anthropic/claude-haiku-4.5,+deepseek-v4-flash,+google/gemini-3.1-flash-lite leaves only those three. Alternatively, type one ID into Enter the name of the custom model (输入自定义模型名称). The ID must match /models exactly, because it is sent unchanged as the model field and an unknown ID is rejected. Current rates are on each model page; translation multiplies them by a lot of tokens, so compare them before you choose.

How many requests a page makes, and staying under rate limits

The OpenAI service sends up to 4 paragraphs per request (Maximum number of paragraphs per request) and, per the official docs, up to 10 requests per second by default, so one long article produces dozens of requests in Dashboard → Logs, each billed for its own input and output tokens. Translations are cached locally for 30 days by default, so reopening a page you already translated may send nothing new. The rate controls sit under Expand for more custom settings: Max requests per second, Max requests per minute, Maximum text length per request and Maximum number of paragraphs per request. Router One applies default request and token limits per key and per account. When a burst crosses one, it answers 429 with RATE_LIMIT_EXCEEDED or TOKEN_QUOTA_EXCEEDED and an X-RateLimit-Scope header naming the limit, and the affected paragraphs fail to translate. Lower Max requests per second (the official docs suggest about 5 for ebooks) or Max requests per minute. APIKEY also accepts several comma-separated keys, which the docs call load balancing, but keys from one Router One account still share that account's limit. If your normal volume needs more, key and account limits can be raised on request: email support@router.one with your account ID, the key's name and your expected peak.

Pick a model for bulk translation, and the quality check

Translation sends the same kind of request thousands of times, so a small, fast, non-reasoning chat model is usually the right trade. Reasonable candidates are anthropic/claude-haiku-4.5, deepseek-v4-flash, deepseek-v4.1-flash, google/gemini-3.1-flash-lite, google/gemini-3.7-flash and openai/gpt-5.6-terra. Avoid IDs that think on every request: anthropic/claude-opus-5.5 always uses adaptive thinking, and IDs ending in -thinking think by default, so each batch also bills thinking tokens at the output rate. Immersive Translate 1.33.3 adjusts some GPT requests itself: for names matching gpt-5.6 and its Sol, Terra and Luna variants, and gpt-5.1, gpt-5.2, gpt-5.4 or gpt-5.5, it sets reasoning_effort to none and drops temperature, and the match also catches prefixed IDs such as openai/gpt-5.6-terra and azure/gpt-5.6-sol. After each answer the extension compares output tokens with input tokens; when the ratio falls outside its range (0.21 to 8 by default), it discards the answer and falls back to another translation. Setting strictPrompt: true skips that check, which the docs recommend only for custom prompts that are not plain translation.

Advanced settings: temperature, timeout, rate and the output cap

Everything above can also be set as JSON under Developer settings → Edit Full User Config, inside translationServices.openai: temperature, requestTimeout in milliseconds (101000 by default for AI services), limit (requests per second), maxTextGroupLengthPerRequest, maxTextLengthPerRequest, bodyConfigs and headerConfigs for extra request fields, modelsOverrides for per-model changes, strictPrompt and langOverrides. The request path lives in baseUrlApiPath, /chat/completions by default, which is why a base URL works. One field deserves attention: 1.33.3 sends its output cap as max_completion_tokens when the host is OpenAI's or Azure's, or when the last segment of the model name starts with gpt-5, o1, o3 or o4, which includes openai/gpt-5.6-terra. Router One's Chat Completions reference documents max_tokens as the output cap, so to cap output, set it in bodyConfigs; the extension then uses max_tokens for every model. Back up your config before editing, since a JSON error makes the extension ignore it:

{
  "translationServices": {
    "openai": {
      "temperature": 0.2,
      "requestTimeout": 60000,
      "limit": 5,
      "bodyConfigs": { "max_tokens": 2048 }
    }
  }
}

Which model ID should Immersive Translate send?

Copy the exact model ID from /models, preserving case, hyphens, and version suffixes; do not substitute a display name. Open its detail page and match the supported API endpoints, context window, and capabilities such as tool calling to the provider and features selected in Immersive Translate. A catalog listing does not mean the client can use every feature of that model. Give each client or application a dedicated API key with a maxSpend cap.

Which API protocol is Immersive Translate using?

OpenAI-compatible describes an interface format; it does not make Chat Completions (/v1/chat/completions), Responses (/v1/responses), and Anthropic Messages (/v1/messages) interchangeable. Check the installed client version, provider configuration, and actual request path against the model detail page and API compatibility fact sheet. A successful plain-text chat does not establish support for hosted tools, conversation state, or file-editing features.

Verify the Immersive Translate call in your request trace

Send a simple text request from Immersive Translate, then match its trace in Dashboard → Logs by time, model, and request_id: tokens, cost, latency, and status. Next, test streaming, tool calls, and multi-turn history separately. For failures, retain the actual request path, full error message, and request_id. If there is no matching log, check client configuration and connectivity before attributing the error to the gateway or upstream.

FAQ