ARK documentation
ARK serves open frontier models behind two APIs you already know: OpenAI's Chat Completions family and Anthropic's Messages. These guides cover every capability, with requests you can paste and responses exactly as ARK returns them.
Quickstart
Get a key and make your first request in two minutes.
API reference
Every endpoint, field, response object and error.
Models
What runs on ARK Cloud, context sizes and capabilities.
Anthropic Messages
Use the Anthropic SDK and Claude-style clients with ARK.
Base URLs
On ARK Cloud, both APIs live on one host and accept the same API key. A self-hosted ARK deployment exposes the same paths on its own host.
| API | Base URL | Use with |
|---|---|---|
| OpenAI-compatible | https://api.ark-labs.cloud/api/v1 | OpenAI SDKs, LangChain, any OpenAI client |
| Anthropic-compatible | https://api.ark-labs.cloud/api/anthropic/v1 | Anthropic SDKs (base URL …/api/anthropic) |
Endpoints
| Method and path | What it does |
|---|---|
POST /api/v1/chat/completions | Chat, reasoning, tool calls, image input, stateful sessions. Guide |
POST /api/v1/embeddings | Dense vectors for search and retrieval. Guide |
POST /api/v1/audio/transcriptions | Speech to text, with subtitles and segment timestamps. Guide |
GET /api/v1/models | The models available right now, with ARK metadata. Guide |
POST /api/anthropic/v1/messages | Anthropic Messages: text, images, tools, thinking. Guide |
POST /api/anthropic/v1/messages/count_tokens | Exact prompt size before you send it. Free. |
GET /api/anthropic/v1/models | The model list in Anthropic's format. |
What is different from OpenAI and Anthropic
ARK aims to behave like the APIs it mirrors, so existing clients work unchanged. A few things are specific to ARK:
- Models are open-weight models with ARK ids such as
arklabsai/Qwen3.6-35B-A3B. Use the exact id from/v1/models; there are no aliases likegpt-4o. - Stateful sessions. Add
ark_statefulandark_max_session_tokensto keep a conversation on the server and send only the new turn. See Stateful sessions. - Reasoning is off by default. Turn it on with
reasoning_effort: "medium"or"high", or Anthropicthinking. The thinking text comes back inmessage.reasoning. See Reasoning. - The default temperature is 0.6, not 1.
- Credits in the usage object. On ARK Cloud, OpenAI-style responses add
credits_usedandcredit_balancetousage. - A few parameters are accepted but have no effect, so clients that always send them keep working:
seed,presence_penalty,frequency_penalty,logit_bias,logprobs,user,metadataand others listed in the API reference. - Every response carries
x-request-id. Quote it when you contact support and we can trace the request.
Where ARK runs
The same API is served by ARK Cloud, our EU-hosted service, and by ARK deployments on your own infrastructure (ARK Core and ARK Tailored). Code written against ARK Cloud runs against a self-hosted deployment by changing the base URL and key.
Getting help
Write to support@ark-labs.pl with the x-request-id of the request in question. For sales and partnerships, use the contact page.