​
Metering & billing

How requests turn into credits, how settlement decides who is paid, and where USDC payout stands.
Every billed request writes one ledger row with two credit amounts: providerCredits and operatorCredits (network revenue). Credits accrue on the ledger today. USDC payout is not enabled. It is planned for public beta and is not active in invite-gated pre-beta.

​
Credit formula

Both divisions use banker’s rounding: ties round to the nearest even value, so small requests carry no systematic bias across large volumes. Constants: If a reported completion count exceeds what the delivered bytes could hold, the billed count is capped at a byte-based estimate.

​
Rate card

The live rate card is a signed feed. Each row sets prompt_rate_per_mtok, prompt_cache_hit_rate_per_mtok, completion_rate_per_mtok, global_multiplier_ppm and provider_share_bps. Rates are credits per million tokens, and usd_per_million_credits is 1.0, so 1,000,000 means $1.00 per million tokens. The default row today: Per-model rows override the default. Every row today has provider share 90% (9000 bps) and multiplier 1.0 (1000000 ppm). Current rows: GET https://api.malibu.tech/v1/rate-card.

​
Faults

Faulted requests produce zero credits:
  • breaker_qualifying: the request faulted. The buyer is not debited and the provider is not credited.
  • null_usage_error: usage was missing, negative, above the overflow guard, or reported more cached tokens than prompt tokens.
A request that never reached a provider (for example 503 no_provider_available) writes no credit row.

​
Prefix-cache billing

Cached prompt tokens are billed at the row’s prompt_cache_hit_rate_per_mtok only on a sticky-conversation hit: the request sent X-MacProvider-Conversation, actually routed back to the provider that served the earlier turn within the sticky TTL (1,800 seconds), and that provider reused the shared prefix. The sticky preference often does not win, so many tagged turns bill at the full prompt rate. The discount applies only on the request’s first routing attempt. A retry on another provider bills at the full prompt rate. See Sticky conversations. Authenticated chat completions also carry an automatic prefix-cache key, so a Mac can reuse its KV cache for a repeated prompt prefix without any header. Those automatic hits are billed at the full prompt rate. The usage object reports the two separately:
  • prompt_tokens_details.cached_tokens (OpenAI-standard) counts every prefix the provider reused, including automatic hits.
  • cached_prompt_tokens (Malibu) counts only the tokens billed at the cache-hit rate.
That example is a sticky second turn. cached_prompt_tokens is 0 when there was no sticky hit. On streaming responses it appears in the final usage chunk. Details: Usage object.

​
Verified model settlement

For covered paid entrypoints, the coordinator’s verdict on the provider’s v0.4 settlement receipt decides whether the buyer is debited and the provider credited. Most requests take this path. A share of requests is dispatched without a settlement snapshot, carries only a v0.3 receipt, and settles as quarantined.
  • verified: the receipt is valid and matches the route-time catalog snapshot. The buyer is debited and the provider credited.
  • pending: the receipt has not arrived or cannot be checked yet. The buyer’s reservation stays held.
  • quarantined: the receipt is missing past the deadline, invalid, from the wrong key, or replayed. Provider credit is zero and the buyer’s reservation is released.
  • Late receipts. A receipt that arrives after the pending deadline (300 seconds by default) is a no-op: no debit and no provider credit.
Full detail: Operations → Verified model settlement.

​
Buyer quota

Buyers do not hold a prepaid balance on the gateway. Each API account has a daily token quota. The gateway reserves tokens when a request starts and settles the reservation against actual usage. GET https://api.malibu.tech/v1/usage returns today’s quota window (daily_tokens_used, daily_tokens_reserved, daily_tokens_remaining, daily_tokens_limit), your API keys, and the settlement disclosure. It does not return credit or dollar amounts.

​
Settlement cadence

  • Today (pre-beta): the coordinator’s weekly settlement job (Monday 00:00 UTC, 7-day window) writes settlement rows on the ledger only. The minimum payout threshold is 500,000 credits, and smaller balances roll into the next window. No USDC transfer runs, because the payout rail is disabled.
  • At public beta (planned): the same weekly windows pay out in USDC on Base once the payout rail is enabled.

​
Endpoints

Regimes: Economics. Provider guide: Payments & payouts.