AI API

AI API: one key for 122 models

Text, images and video through one OpenAI-compatible API: 49 text models, 37 image models and 35 video models. GPT, Claude, Gemini, Grok, DeepSeek, Midjourney, Veo and Kling are all called the same way, with one key and one balance.

PythonJavaScriptcURL

from openai import OpenAI

client = OpenAI(
    base_url="https://api.altrouter.ai/v1",
    api_key="ar-...",
)

resp = client.chat.completions.create(
    model="gpt-5.6",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(resp.choices[0].message.content)

Create an account and get a key Documentation

Three kinds of requests

Everything goes to the same base URL, https://api.altrouter.ai/v1, with the same key. Only the endpoint changes.

What	Endpoint	How it works
Text and chat	`POST /chat/completions`	whole response or streamed (SSE), with tool calling, structured outputs and image input
Images	`POST /images/generations`	generation and reference-based editing; we keep the file indefinitely
Video	`POST /videos`	asynchronous: create a job, poll its status, download the clip
Anthropic format	`POST /messages`	for Claude Code and the Anthropic SDK: same catalog, same balance
Catalog and balance	`GET /models`	models, prices and parameters; remaining balance via GET /credits

Image

Python

from openai import OpenAI

client = OpenAI(base_url="https://api.altrouter.ai/v1", api_key="ar-...")

img = client.images.generate(
    model="nano-banana-2.1",
    prompt="neon Tokyo in the rain",
    size="1024x1024",
)
print(img.data[0].url)

Video

A clip takes minutes to render, longer than any HTTP request should stay open, so video works as a job: you create it, poll it and pick up the result. Details are in the video docs.

Python

import requests, time

h = {"Authorization": "Bearer ar-..."}
B = "https://api.altrouter.ai/v1"

job = requests.post(f"{B}/videos", headers=h, json={
    "model": "veo-3.1",
    "prompt": "neon Tokyo in the rain",
    "seconds": 8,
}).json()

while job["status"] not in ("completed", "failed"):
    time.sleep(10)
    job = requests.get(f"{B}/videos/{job['id']}", headers=h).json()

How to get a key

Sign up with Telegram, Google or an email link. No password needed.
Confirm your Telegram and $1 lands on your balance for testing, once per person.
Create a key in the dashboard; it starts with ar-. The key is shown only once, so save it right away.
Top up your balance when the test dollar runs low: Russian bank card, SBP (Russia’s Faster Payments System) or crypto, packages from $15.

Create an account and get a key

Pricing

You pay per request, there is no subscription. The vendor’s official price is shown struck through next to ours: text models are 10–29% cheaper, and across the catalog the discount goes up to 64%.

Model	Official price	Our price	Discount
GPT-6 Astra	$10.00 / $50.00	$8.50 / $42.50	−15%
GPT-6.1 Sol	$2.00 / $10.00	$1.70 / $8.50	−15%
Claude Opus 5.5	$4.00 / $20.00	$3.40 / $17.00	−15%
Claude Sonnet 5.5	$2.00 / $10.00	$1.70 / $8.50	−15%
GPT-5.6 Terra	$2.00 / $12.00	$1.70 / $10.20	−15%
Gemini 3.1 Pro	$2.00 / $12.00	$1.46 / $10.20	−15%
Gemini 3.8 Flash	$0.75 / $3.75	$0.64 / $3.19	−15%
Grok 4.7	$2.00 / $6.00	$1.70 / $5.10	−15%

Text: USD per 1M tokens, input / output.

Model	Official price	Our price	Discount
GPT Image 2.5	-	$0.042 per image	-
Qwen Image 3.0	$0.030	$0.028 per image	−8%
Nano Banana 2.1	$0.034	$0.024 per image	−30%
Veo 3.1	$3.20	$2.55 per clip	−20%
Seedance 2.5	$0.47	$0.39 per sec	−17%
Hailuo 2.3	$0.047	$0.037 per sec	−20%

Images: per image; video: per second or per clip.

The full catalog: 122 models with prices and parameters.

What the API can do

Streaming of the response as it is generated, in the OpenAI format.
Tool calling and structured outputs with JSON Schema.
Image input for vision models, and reasoning models.
Fallback provider: on an error or timeout the request goes to another provider, where a fallback exists.
The cost of every request comes back in the response, and the full log is kept for a year.
Per-key limits: a spending cap per day, per month or in total, and a requests-per-minute limit.

What it works with

The official OpenAI SDKs for Python and JavaScript, and any client that lets you set your own base_url, such as Cherry Studio and Dify. Moving over from OpenRouter means swapping the URL and the key; see the OpenRouter alternative page for details.

Claude Code

Claude Code talks to an Anthropic-compatible /v1/messages endpoint, and we have one. Three environment variables are all it takes:

Shell

export ANTHROPIC_BASE_URL=https://api.altrouter.ai
export ANTHROPIC_AUTH_TOKEN=ar-...
export ANTHROPIC_MODEL=claude-sonnet-5.5
claude

What we don’t have: embeddings, reranking, TTS/STT and batch mode.

FAQ

What is an AI API?

It is a way to call an AI model from your own code: your program sends a request to the API endpoint and gets back the model’s reply, an image or a video. Through AltRouter every model is called the same way, in the OpenAI format, with one key.

API documentation

How do I get an AI API key?

Sign up with Telegram, Google or an email link and create a key in the dashboard. Confirming your Telegram adds $1 to your balance, which is enough for your first requests. One key works with every model in the catalog.

How to get an AI API key

Which AI models are available through the API?

122 models, including 49 text models (GPT, Claude, Gemini, Grok, DeepSeek, Kimi and others), 37 image models (Nano Banana, GPT Image, FLUX, Midjourney, Seedream) and 35 video models (Veo, Kling, Seedance, Hailuo).

Vendors and their models

How much does an AI API cost?

You pay per request, with no subscription. Text is billed per million input and output tokens, images per image, video per second or per clip. Most models cost less than the vendor’s official price; for text models the difference is 10–29%.

Model catalog and prices

Can I pay for an AI API from Russia?

Yes. You top up your balance with a Russian bank card, via SBP or with crypto, in packages from $15. Rubles are converted to dollars at the USDT/RUB exchange rate, and you see the ruble amount before you pay.

AI aggregators you can pay for from Russia

Do I need a separate key for each AI model?

No. One key and one balance cover every model; you choose the model with the model field in the request. Separate keys make sense for different services or team members, so you can see each one’s spending and give each its own limit.

API limits

Which tools does the API work with?

With the official OpenAI SDKs for Python and JavaScript, with any client that lets you set your own base_url (Cherry Studio, Dify and others), and, through the Anthropic-compatible /v1/messages, with Claude Code and the Anthropic SDK.

Connecting Claude Code from Russia

Get connected
in two minutes.

A free key, $1 in credits and every model at once - no card required.

Create an account Read the docs