Developers / Docs / Usage and billing

Usage and billing

ARK Cloud is prepaid and metered per request. Every response tells you what it used and what it cost.

Credits

You buy credits on the ARK Cloud portal: one million credits cost $1. Each request deducts its cost from your balance; when the balance reaches zero, requests return 402 until you top up. Prices per model are on the pricing page.

What is billed

Request	Billed as
Chat and messages	`prompt_tokens` at the input price plus `completion_tokens` at the output price. The prompt includes the chat template, system prompt and tool definitions; the completion includes reasoning tokens.
Images	Image tokens are part of the prompt. Vision models add a small per-image-token surcharge.
Stateful turns	Only the tokens the turn sends and generates. The stored conversation is not billed again.
Embeddings	Inputs × the length of the longest input in the batch, at the input price.
Transcription	The duration of the audio file, silence included, at the price per audio hour.
Free	`count_tokens`, model listings, and requests that fail before producing output.

Each request's cost is rounded up to a whole credit. Cached prompt tokens (cached_tokens) make responses faster but are billed like any other prompt token.

Reading usage

usage on /api/v1/chat/completions

"usage": {
  "prompt_tokens": 1820,
  "completion_tokens": 412,
  "total_tokens": 2232,
  "prompt_tokens_details": { "cached_tokens": 1536, "audio_tokens": 0 },
  "completion_tokens_details": { "reasoning_tokens": 260, "audio_tokens": 0, "accepted_prediction_tokens": 0, "rejected_prediction_tokens": 0 },
  "credits_used": 214,
  "credit_balance": 4783120
}

credits_used is what this request cost; credit_balance is what is left.
In a stream, usage arrives in the final chunk when you set stream_options.include_usage.
The Anthropic route reports tokens in its own format (input_tokens + cache_read_input_tokens, output_tokens) and bills the same way.

When a request is charged

A request is charged when it completes: after generation for normal responses, and at the end of the stream for streamed ones. Requests rejected with an error are not charged.

A worked example

The usage above, on arklabsai/Qwen3.6-35B-A3B at $0.0156 per million input tokens and $0.4499 per million output tokens:

Cost

input   1,820 tokens × 0.0156 credits = 28.39 credits
output    412 tokens × 0.4499 credits = 185.36 credits
total                                  213.75 → 214 credits ($0.000214)

One credit per token per dollar-per-million: a price of $0.4499 per million tokens is 0.4499 credits per token.