Skip to main content
Darkbloom has two paid products with separate pricing. Inference is charged per token from a prepaid credit balance. Darkbloom Sandbox CI jobs are charged per second of runner time, with a one-minute minimum. There are no subscription or seat fees.

Inference pricing

Inference rates are per token and vary by model. The catalog and its prices are maintained by the coordinator, so fetch the current values instead of relying on a fixed table:
GET /v1/models repeats the same prices in each model’s pricing block (prompt, completion, input_cache_read, image, request). The request path is described in Billing. Prices are µUSD per 1,000,000 tokens; the *_usd fields are the same rates as USD per 1M tokens. This snapshot was read from the live endpoint on 2026-11-16: If a model has no price row, the fallback rates are $0.0500 input, $0.2000 output and $0.0250 cache read per 1M tokens. During the public alpha the platform fee is 0%, so providers keep 100% of the per-token revenue. A provider may set its own price only at or below the platform price, so a request never bills a token above the rate GET /v1/pricing shows.

How inference billing works

  1. Before dispatch, the coordinator reserves the worst-case cost: the estimated prompt plus the full output bound (your max_tokens, or the model’s default when you omit it).
  2. When the request settles, the coordinator charges the actual token counts and refunds the unused part of the reservation.
  3. If the balance cannot cover the reservation, the request is refused with 402 (insufficient_funds), not 403. A per-key spend cap has its own 402 (insufficient_quota).
Prompt tokens served from a provider’s prefix cache bill at the model’s cache-read rate and are reported as usage.prompt_tokens_details.cached_tokens.

Darkbloom Sandbox pricing

Sandbox CI jobs are priced per runner class and minute, then billed per second:
  • A job holds the price of 60 minutes of its runner class when it starts.
  • It settles on the seconds it actually ran, with a one-minute minimum and a 60-minute maximum.
  • The current price sheet keeps a 10% fee; the provider receives 90% of the job’s time charge.
  • Jobs that fail for infrastructure reasons are released and cost nothing.
  • Jobs that run on the payer’s own provider Mac are free.
  • A job keeps the price in effect when it started, even if prices change while it runs.
Signed-in accounts read the current runner-class rates and fee from GET /v1/sandbox/rates. Before Darkbloom publishes a price sheet, that route returns 503 sandbox_billing_unavailable and no job can start. The formulas and constants are in the coordinator’s pricing reference; the customer-facing consent and charge routes are in Add credits.

Check your balance

The response includes your spendable balance and the withdrawable amount (provider earnings and referral rewards):
balance_micro_usd is the canonical integer value. withdrawable_micro_usd is the part you earned and can withdraw; deposits and invite credits are spendable but not withdrawable.

Check usage history

Each entry in usage is a settled request:
cached_tokens is the part of prompt_tokens billed at the cache-read rate and is omitted when zero. cost_micro_usd is the cost in millionths of a dollar. Usage is a recent-history view, not a complete billing export.