Compute Portal

Pricing model ​

Compute is billed one month in advance, auto-renews monthly, and can be cancelled before the next renewal date. There is no per-second, per-minute, or per-hour billing on compute plans.

This page is the authoritative description of how Compute Portal bills. Every other doc — API reference, CLI help, MCP tool descriptions, llms.txt — defers to this page.

Compute plans ​

The core billing unit is the monthly compute plan (vCPU + memory + storage, with optional GPU). You pick a SKU from the catalog, authorize the first month, and the plan renews monthly at the then-current rate.

  • Charged: full month, up front, when you confirm the order.
  • Renewal: automatic, same day each month.
  • Cancel: anytime before the renewal date. The current period runs to its end; there is no partial-period refund.
  • Toggle auto-renew: cpctl billing auto-renew <subscription-id> --off keeps the plan until the renewal date without charging again.

GPU jobs and reservations ​

cpctl gpu run and cpctl gpu reserve draw from your monthly plan, not a separate meter.

  • A one-shot GPU job uses capacity your plan has already paid for. The job runs, exits, and releases the capacity back to your plan. No per-second charge is applied.
  • A GPU reservation holds a slice of your plan's capacity for a window. If your plan does not cover the requested GPU hours, the reservation fails before any work starts.
  • The cost_per_hour field in API responses is a projection of your monthly plan rate divided across the reservation window. It is a reference number, not a usage meter — you are not charged hourly.

Agent-key spend ​

This is the one place real per-call billing exists, and it is not the same product as compute plans.

  • Scoped API keys (cp_key_*) and cpctl pay authorize tokens are micro-metered per call against a hard dollar cap you set when the key is minted.
  • Spend applies only to agent-driven calls using that scoped key. It does not affect your compute plan.
  • See cpctl pay authorize for the pattern.

Common misreadings ​

You might seeWhat it actually means
cost_per_hour in a reserve responseA projection of your monthly plan rate. You are not charged separately.
"Billed per hour" in a reference-rate tableReference projection. Compute itself is still prepaid monthly.
An old doc saying "per second"Stale copy. The policy on this page is authoritative. File an issue if you find one that we missed.

GPU Compute Platform

© 2026 Compute Portal