Pricing & fees

Where the hosted number comes from

The Hosted column is a published output price, not our estimate. We anchor to the nearest listed model on Ollama's price sheet, cross-checked against DeepSeek's, whose v4-flash tier lists $1.32 per million output tokens at peak and half that off-peak.

Read those sheets and the first surprising thing is how flat they are. An 8B and a 70B are separated by a factor of roughly two and a half, not by the factor of nine their parameter counts suggest. Managed providers batch aggressively: the weights are read from memory once per forward pass and produce a token for every request in flight, so a large model's cost is spread across dozens of concurrent streams. The book prices against those numbers, not against a markup we picked.

Escrow

Posting a bid places a hold of size × limit on your funded balance, plus the 2% buyer fee on that amount, so the hold covers the worst case the order could cost you. It is not a charge and it is not spent. Cancelling releases it immediately, and a partial fill releases the unfilled remainder.

Because a bid fills at the ask rather than at its own limit, the hold is almost always larger than the eventual charge. That difference comes back to you.

What a job actually costs

Output tokens at the price the order filled at, plus prompt tokens at 25% of that same price. Prefill runs in one parallel pass instead of a token at a time, so it costs a host a fraction of what generating the same count does, and the bill follows the work: a quarter, not nothing and not full price.

Send the next turn back on the same order and it can land on the node that just served you while that node still holds the conversation in cache — for 5 minutes after the last turn settles. The part of your new prompt the node already holds is read from cache instead of prefilled again, and it bills at 2.5% of the fill price — a tenth of the fresh prompt rate, the same discount managed providers publish for a warm read. Only the overlap is discounted; anything new in the prompt is fresh input at the 25% rate. A turn that lands on a different node, or arrives after the window, is a clean miss and bills the whole prompt fresh — you are never charged the cache rate for a read that did not happen. Keeping an agent's loop on one order is what lets it earn that discount; a fresh bid each turn starts cold.

Your prompt is charged once per request, at the price that request filled at. If a bid crosses several asks and the answer is stitched across more than one node, the prompt is still charged once — being served by two machines instead of one is the venue's business, not something you pay extra for.

Both counts are the venue's, produced by tokenising the prompt it sent and the response it received. The node's own report is recorded for comparison but never billed from — see verification. That matters more for input than for output: a host that could report its own billable prompt size could invent one.

The hold a bid places covers the prompt as well as the reply, so an order is never accepted that cannot pay for the prompt it carries.

You pay for output whether or not the completion ran to the end. A stream you abort, one the serving node drops halfway, and one cut short at its token ceiling are all billed for exactly the tokens that reached you and refunded the rest of the hold — a short answer is never a full charge. Which of those happened is on finish_reason; the API reference spells out each and when a follow-up request is worth sending.

Payouts

What a job earns is credited to your venue balance the moment it settles, and you can see it on your account immediately. Once your earnings reach $50.00 you can request a payout to a Bitcoin address you nominate, from the host page or the funding page. The same threshold covers host earnings and referral commission. Below it the network fee would eat most of the payout, which is the only reason a floor exists.

Only what you earn here can be paid out. Money you deposited buys compute and stays on the venue: taking coin in at one address and paying it out at another is a money transmission business, and this is a compute market instead. So the payout figure on your account is your fills and referral commission net of the venue’s cut, less anything already requested — not your whole balance.

Requesting debits your balance straight away, so the money is out of reach while the request is open and cannot be spent or claimed twice. The coins are sent afterwards: every payout is reviewed and signed by a person, offline. Nothing about moving real Bitcoin is automated here, and that side is still being built — so treat the turnaround as manual and unhurried rather than same-day. If a request is rejected the balance returns in full as a visible credit on your ledger, not a silent reversal.

There is no bond and nothing is staked or slashed. Earlier versions of this page said otherwise; that was wrong and has been removed rather than quietly softened.

How the fee is split, and what it pays for

The venue charges 2% to the buyer and 3% to the provider, booked as two separate ledger rows so each side can see exactly what it paid. The weighting is deliberate: buyers have an obvious alternative in the hosted API they are already using, so their side is kept as thin as it can be while still being real. Providers are selling capacity that would otherwise earn nothing at all, and 97% of something beats 100% of an idle card.

What the house collects is not margin skimmed off a match and forgotten. It funds the things that make the venue worth routing through in the first place: keeping the exchange and its settlement path online, paying people to investigate disputed jobs and failed audits, and staffing support so a host with a node that will not register or a buyer with a job that stalled can talk to a human.

The fee is a transfer, not a burn. Every micro-USD taken off either side is credited to the house account inside the same ledger, which is why the fee column sums to zero across the whole book and the venue's revenue is a balance you can audit rather than a number reconstructed by subtraction.