What you pay, and why it moves

Prices are in euros per million tokens, listed per model, with input and output priced separately. No subscription, no minimum, no negotiation for the first million.

Where the price comes from

Two layers. Every model ships with a price in the build; an operator can override it from the dashboard, and the change applies to the next job — no deploy, no release. Pricing is a business decision, and a business decision that needs an engineer is a business decision that happens too slowly.

The guard rails are enforced on write: a model that does not exist cannot be priced, output can never be cheaper than input, and the promised context can never exceed what the least capable host actually serves. You cannot be routed to a machine that silently gives you less than the page said.

What the 80% means for a host

Of every euro a job earns, 80 cents go to the machine that produced the tokens. In practice:

host €/hour  =  measured tokens/second × 3600 ÷ 1e6 × output €/Mtok × 0.8

Your dashboard runs exactly that formula, using the decode rate your own card measured — never a projection, and never a monthly figure, because a monthly figure would require promising demand we cannot promise. See what your GPU actually earns.

Keeping a bug from costing you money

• Prepaid balance. You cannot spend money you have not added. An empty balance returns 402, not an invoice.
• Per-key spend caps. Every key carries one (€20 by default). A runaway loop hits the cap and stops, rather than discovering your credit limit.
• Usage per key. The dashboard breaks spend down by key and model, so "which integration is expensive" is answerable.

Why we are cheaper

The compute is hardware that already exists and is otherwise idle: gaming rigs, workstations, homelab boxes across Europe. We do not amortise datacenter construction into your per-token price. That is also the honest limit — capacity depends on how many hosts are online, which is why the batch and interactive tiers are separated and why we publish measured throughput rather than best-case numbers.