Pricing & unit economics

The catalog lists real per-million-token input/output prices for every published model, currently OpenRouter upstream cost plus a 5% markup. Charges are computed in integer microdollars from actual reported usage and capped by the request reservation. Old prototype prices, charts, spreads, and savings counters were never production pricing.

Text inference often charges input and output tokens at different rates. Some models also have cache, reasoning, context-length, request, or other pricing conditions. GPU runtime pricing is a different cost model: the operator can pay for startup and idle time even when no billable output reaches a buyer.

base charge =
  (input tokens × input rate per million
   + output tokens × output rate per million) / 1,000,000

customer charge = base charge + disclosed applicable fees

A production quote must specify its units, supported workload, ceiling, expiry, and fee policy. Estimates before generation are not final output costs. Amounts used for reservations and settlement should use exact integer or decimal arithmetic with deliberate rounding rules.

• Measure actual supplier or GPU cost, including startup, idle time, failed attempts, and utilization.
• Include payment/facilitator costs, gas where applicable, infrastructure, support, fraud, and refunds.
• Distinguish direct gross margin from net business profit.
• Do not rely on token trading volume or speculative market capitalization to prove demand for inference.
• Use representative quality and latency targets; cheaper output that fails the task is not equivalent value.

Supplier competition can improve value, but it does not automatically beat a large provider's batching efficiency or wholesale rates. If a self-hosted model fails cost or reliability targets, an authorized wholesale arrangement may be the better supply path.

‘Up to 90% cheaper’ is not a universal product guarantee. Any published saving needs a named baseline, equivalent model/service assumptions, workload details, measured cost, and limitations. Input reduction is not the same as total bill reduction, particularly when output cost dominates.

Graph-based retrieval or other context-selection tools may help repeated, context-heavy tasks. They are not an automatic discount on every prompt. Free API quotas are useful for permitted experiments, but they are not proof of resale rights, privacy guarantees, production capacity, or sustainable margins.