Documentation

1Quick start

1. Create an account

   Open the sign-up page and enter a username, password and email address. Click “Send code”, enter the code from the email, agree to the User Agreement and Privacy Policy, then click “Create account”. After that, sign in with your username or email.

   

2. Add funds

   New accounts currently start with US$0.05 of credit, and the key they come with is already on “Lowest price”. To make it last, start with a smaller model; see the setup guides.

   After signing in, open the console and click “Wallet” in the left menu. Under “Add Funds”, choose an amount and a payment method. Your balance is in USD and you get exactly what you pay; it usually arrives within a few seconds of payment. Payment methods and limits are under Top-ups and refunds.

   

   

3. Get your API key

   New accounts come with a key called “Default Key”, on “Lowest price” (the cheapest groups, sometimes less stable). Click “API Keys” in the console’s left menu to find it. Once you have topped up, you can switch it to “Auto” by editing it.

   In the “API Key” column of the list, click the copy icon. This copies the full key (the list only shows it masked with asterisks). You can also click “⋯” at the right end of the row and choose “Copy Key”.

   

   Keys start with sk-; requests work with or without this prefix. A key gives access to your balance, so never put it in front-end code or a public repository. If a key leaks, delete it under “API Keys” and create a new one.

   To make another key, click “Create API Key”. Enter a name, keep “Smart routing” on (a new key starts on “Auto”, or on “Lowest price” if your balance is under US$1; we rank the groups for you, and retry another group if one fails), pick another mode, or turn it off and choose groups under “Group priority” (see Models and groups), and click “Save changes”. If needed, you can set an expiration time; turn off “Unlimited Quota” to give the key a spending cap; under “Advanced Settings” you can limit which models it can use and set an IP whitelist (CIDR supported).

   

4. Set the base URL

   Use the address that matches the API format your code or client uses:

   API format	Base URL	Send the key in
   OpenAI-compatible	https://api.steadyrouter.com/v1	Authorization: Bearer
   Anthropic native	https://api.steadyrouter.com	x-api-key or Authorization: Bearer
   Gemini native	https://api.steadyrouter.com	x-goog-api-key or ?key=

   api.steadyrouter.com only serves the API, so opening it in a browser shows 404. That is expected.

5. Send your first request

   Store your key in the environment variable STEADYROUTER_API_KEY, then send a request in any of these ways:

   curlPythonNode.jsCopy

   export STEADYROUTER_API_KEY="sk-..."
   
   curl https://api.steadyrouter.com/v1/chat/completions \
     -H "Authorization: Bearer $STEADYROUTER_API_KEY" \
     -H "Content-Type: application/json" \
     -d '{
       "model": "gpt-5",
       "messages": [{"role": "user", "content": "Hello"}]
     }'

   To see which models your key can use, request the model list:

   curlCopy

   curl https://api.steadyrouter.com/v1/models \
     -H "Authorization: Bearer $STEADYROUTER_API_KEY"

2APIs

The same key works with all three formats; use whichever your code or client already speaks. Request and response parameters match the official APIs; see the official documentation for parameter details.

OpenAI-compatible

Base URL https://api.steadyrouter.com/v1, key in Authorization: Bearer. Official docs: OpenAI API Reference.

Not available: file uploads (Files), Batch, Fine-tuning, Assistants.

Anthropic native (Messages)

Base URL https://api.steadyrouter.com (the SDK appends /v1/messages itself). Send the key in x-api-key or Authorization: Bearer; the anthropic-version header is optional. Official docs: Messages API.

• POST /v1/messages, with streaming. The request body is passed upstream as is, including fields such as cache_control and thinking.
• /v1/messages/count_tokens is not available (returns 404). Claude Code falls back to an estimate when this happens, so it keeps working.
• For Anthropic-specific features such as prompt caching and extended thinking, use this native format; calling Claude in the OpenAI format is only suitable for plain chat.

curl https://api.steadyrouter.com/v1/messages \
  -H "x-api-key: $STEADYROUTER_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-5-5",
    "max_tokens": 1024,
    "messages": [{"role": "user", "content": "Hello"}]
  }'

Gemini native

Send the key in the x-goog-api-key header or in the ?key= URL parameter. Official docs: Gemini API.

curl "https://api.steadyrouter.com/v1beta/models/gemini-2.5-flash:generateContent" \
  -H "x-goog-api-key: $STEADYROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents": [{"role": "user", "parts": [{"text": "Hello"}]}]}'

Midjourney and Suno

Midjourney uses the endpoints starting with /mj/ (midjourney-proxy protocol, such as /mj/submit/imagine, /mj/submit/action and /mj/task/{id}/fetch); Suno uses /suno/submit/music and /suno/fetch/{id}. For both, the base URL is https://api.steadyrouter.com and the key goes in Authorization: Bearer. Both are billed per request.

3Clients

Any SDK or app that lets you set a custom base URL works.

Step-by-step setup guides: Cherry Studio · Cline · Dify · Kilo Code · LobeHub · n8n · Open WebUI · opencode · WorkBuddy

OpenAI SDK

Change base_url to https://api.steadyrouter.com/v1 and set api_key to your key; the rest of your code stays the same. See step 5 of Quick start for an example.

Anthropic SDK

# pip install anthropic
import os
import anthropic

client = anthropic.Anthropic(
    base_url="https://api.steadyrouter.com",
    api_key=os.environ["STEADYROUTER_API_KEY"],
)
msg = client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello"}],
)
print(msg.content[0].text)

// npm install @anthropic-ai/sdk
import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({
  baseURL: "https://api.steadyrouter.com",
  apiKey: process.env.STEADYROUTER_API_KEY,
});
const msg = await client.messages.create({
  model: "claude-sonnet-5-5",
  max_tokens: 1024,
  messages: [{ role: "user", content: "Hello" }],
});
console.log(msg.content[0].text);

Claude Code

Set two environment variables: ANTHROPIC_BASE_URL to the address and ANTHROPIC_AUTH_TOKEN to your key.

export ANTHROPIC_BASE_URL="https://api.steadyrouter.com"
export ANTHROPIC_AUTH_TOKEN="sk-..."
claude

To make them permanent, put them in the env block of ~/.claude/settings.json. We also recommend pinning the three default model tiers to models we offer: Claude Code’s built-in defaults are not necessarily on our model list, and through a gateway, background tasks such as session titles use the main model by default, so pinning a cheaper Haiku model saves money.

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.steadyrouter.com",
    "ANTHROPIC_AUTH_TOKEN": "sk-...",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-opus-5-5",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-5-5",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-haiku-4-5-20251001"
  }
}

• Which group serves Claude Code is set on the key, not here: keep smart routing on for the key, or turn it off and choose groups for the key under “API Keys” (see Models and groups). The model names above stay the same.
• After starting, run /status; if it shows the base URL https://api.steadyrouter.com, you are set up.
• If requests fail with 400 and the error mentions context_management or Extra inputs are not permitted, also set the environment variable CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1.
• Claude Code’s official guide: Connect Claude Code to an LLM gateway.

Codex CLI

Add a provider in ~/.codex/config.toml and put your key in the environment variable STEADYROUTER_API_KEY. Codex uses the Responses API (wire_api = "responses"). For the model name, use one of the Codex models on Model Square.

model = "YOUR_CODEX_MODEL"
model_provider = "steadyrouter"

[model_providers.steadyrouter]
name = "SteadyRouter"
base_url = "https://api.steadyrouter.com/v1"
env_key = "STEADYROUTER_API_KEY"
wire_api = "responses"

Cherry Studio and other desktop clients

In the app’s settings, add a model provider of the OpenAI type, enter your key as the API key and the address according to the general rule above, then fetch the model list or add models by hand. For Claude models you can also add a provider of the Anthropic type with the address https://api.steadyrouter.com.

4Models and groups

All models, prices and billing methods are on Model Square, so we don’t repeat them here. In code, GET /v1/models returns the models your key can use.

Groups: one model, several sources

A model can be offered in several groups. Each group is a different upstream source with its own price, feature support, speed and availability. Group names read product-source-number, for example Claude-Bedrock-01; the source part tells you the type:

• Which groups serve your requests is set on your API key: model names in your code stay the same whichever group serves a request, and a request can also ask for a routing mode (see “Choosing the mode per request” below). Model Square shows each group's price and its availability over the last 24 hours.
• The key a new account comes with has “Smart routing” on, in the “Lowest price” mode; keys you create have it on too, in the “Auto” mode (or “Lowest price” if your balance is under US$1). With smart routing on, the gateway ranks the groups that offer the model and tries them in that order: if a group fails (other than a 408, 504 or 524 timeout), the request is retried in the next group within the same call, up to 5 groups. A 400 error from upstream is retried in at most one more group, at another upstream site; if that fails too, you get the first error back. “Auto” puts cloud and API groups first, CN, Mixed and image groups next and Pool groups last, and within each of these ranks by recent success rate, price and speed. “Lowest price”, “Fastest” and “Most reliable” rank all groups by that one measure, so a Pool group can come first. In every mode, a group that failed at least half of 3 or more attempts in the last 30 minutes moves to the end: it is tried only if the others fail.
• To choose the groups yourself, turn smart routing off and add groups under “Group priority”: requests use only those groups, in your order, and move to the next one when a group fails. With a single group, requests go to that group only, and the error is returned when it fails. If a group does not suit you, replace it on your key.
• The Failover tag on Model Square means the model is offered in groups from at least two independent suppliers. A request moves to another group when one fails only if smart routing is on or the key has two or more groups under “Group priority”; a key with a single group does not switch.
• Groups differ in feature support (for example prompt caching), speed and output quality, and which group you use is your choice. For stability-critical work, use only cloud or API groups: turn smart routing off and list only such groups under “Group priority” (in every smart routing mode, Pool and Mixed groups can be used).
• Every day we check, from real traffic, that each group returns the model you requested.
• The model field in the response is the name returned upstream and may include a date version. That is normal; billing uses the model name you requested.
• Prices change, up or down, when our costs from upstream suppliers change, sometimes several times a day. Each request is charged at the price in effect when it is made, for the group that served it.

Choosing the mode per request

A request can ask for a smart routing mode itself, which overrides the key's mode; :floor, :nitro and provider.sort are written the same way as on OpenRouter.

• Model name suffix: :floor or :price for “Lowest price”, :nitro or :speed for “Fastest”, :stable or :success_rate for “Most reliable”, for example "model": "gpt-5:floor".
• Or in the request body: "provider": {"sort": "price"}, with price, speed (also throughput or latency) or success_rate. If both are given, the suffix wins. Other provider options are not supported.
• With smart routing on, this reorders the same groups; with your own group list, it reorders only the groups in your list. The gateway removes the suffix and the provider field first, so billing and the upstream see the plain model name. It works for JSON request bodies: image edits sent as multipart form data do not accept a suffix (the request fails as an unknown model), and on Gemini-native paths, where the model is in the URL, only provider.sort counts.

5Billing

Prepaid, with the balance in USD. Each request is charged from your balance at the price listed on Model Square at that time for the group that served it, in one of two ways:

How caching is billed

Cache reads and cache writes are billed at the cache read and cache write prices, based on the token counts in the response’s usage; the rest of the input is billed at the normal input price. The fields in each format:

• OpenAI format: usage.prompt_tokens_details.cached_tokens
• Anthropic format: usage.cache_read_input_tokens (read), usage.cache_creation_input_tokens (write)
• Gemini format: usageMetadata.cachedContentTokenCount

When you are not charged

• Requests that return an error are not charged, except that some upstream rejections for breaking content rules carry a violation fee.
• When the gateway retries a request in another group, the attempts that failed and were retried are not charged.

Checking usage and controlling spend

• Console → “Usage Logs” shows the model, group, token counts, cost and request ID of every request.
• Once your balance runs out, requests return 403 (insufficient_user_quota); they work again as soon as you top up.
• A spending cap on a key keeps any one program from using up your balance.

6Rate limits and errors

Rate limits

• Each account can send at most 300 requests per minute, failed requests included. All keys in the same account share this limit. Requests over it return 429.
• The model list (GET /v1/models), Midjourney and Suno endpoints do not count.
• If you need higher concurrency, email support@steadyrouter.com.

Error format

Errors generated by the gateway itself use the format below (on the Anthropic and Gemini endpoints too). Some error messages are in Chinese, so have your code check the HTTP status code and error.code.

{
  "error": {
    "code": "model_not_found",
    "message": "分组 auto 下模型 no-such-model 无可用渠道（distributor） (request id: 20261008...)",
    "type": "new_api_error"
  }
}

Every response carries the response header X-Oneapi-Request-Id, and error messages end with (request id: …). Include it when you contact us so we can find that request.

Status codes and whether to retry

Timeouts

We do not limit the total duration of a request; a streaming request is closed after 600 seconds without any data. Non-streaming requests such as long reasoning or image generation can take a while, so set your client timeout to 300 seconds or more; use streaming for requests with long output where you can.

7Image generation

Which endpoint

• Most image models use POST /v1/images/generations (text to image); image editing uses POST /v1/images/edits (a multipart form that uploads the source image).
• Gemini models with image in their name are called through the chat endpoint POST /v1/chat/completions.
• For Midjourney (mj_*), see APIs.

curl https://api.steadyrouter.com/v1/images/generations \
  -H "Authorization: Bearer $STEADYROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  --max-time 300 \
  -d '{
    "model": "doubao-seedream-3-0-t2i-250415",
    "prompt": "a red apple on a white table, minimal product photo",
    "size": "1024x1024",
    "n": 1
  }'

Size and quality

Pass size, quality, n and other parameters as the model vendor defines them; the gateway forwards them upstream as is. The vendor’s documentation decides which sizes and quality levels each model supports, for example OpenAI’s Images API.

The resolution and quality an image model actually delivers can differ between groups. If a group does not give you what you need, switch your key to another group.

Response format

Some models return b64_json (the image in base64), others return url (a temporary link that expires, so download and save it promptly). Handle both.

Duration and timeouts

The image endpoints are not streaming: they return once the whole image is generated. A single image usually takes tens of seconds; the longest we have measured is about 135 seconds. Set your client timeout to at least 180 seconds, ideally 300; we will not close the connection early.

Billing

Per-request models have one price per image, regardless of size and quality. For per-token models (such as the gpt-image series), larger sizes and higher quality mean more output tokens and a higher cost. Failed generations are not charged, except that some upstream rejections for breaking content rules carry a violation fee.

8Top-ups and refunds

Top up under “Add Funds” on the console’s “Wallet” page, in whole US dollars. There are no fees: your balance goes up by exactly the amount you pay.

• During the first 7 days after sign-up, a new account can top up at most US$50 per day in total by card, Apple Pay or Google Pay. WeChat Pay is not subject to this limit.
• Paying may open a new tab, window or app. If you are not taken back automatically after paying, switch back to the payment page and make sure the payment is complete.
• Your balance does not expire. It can only be used on this site and cannot be withdrawn or transferred to another account.

Refunds

• Unused balance: within 7 days of a top-up, you can ask for a refund of the part of it you have not used. Usage is counted against your oldest top-up first. The refund goes to the original payment method, minus the payment processing fee on that amount and the US$1 fee the payment provider charges for each refund; if nothing is left after these fees, no refund is issued.
• Used balance is not refundable. Credit received for free, through promotions or from redemption codes is not refundable.
• Billing errors: send us the request ID of that request; once we confirm the error, we credit the full amount back to your balance.
• Refunds can only be requested by email: write to support@steadyrouter.com from the email address on your account, or include your username. See section 3 of the Terms of Service for the full terms.

9Contact us

For accounts, billing, refunds, billing errors or reporting abuse, email support@steadyrouter.com.

• Write from the email address on your account, or include your username.
• For a specific request, include its request ID, time and model.
• Never send us your API key; we will never ask you for one.