One prepaid wallet pays for every backend (1 credit = €0.01 — see Billing). Simulators are free; QPU shots and GPU/CPU minutes draw from the wallet with a reserve-then-settle model that makes overspending impossible. How prices are set. Paid compute is priced from cloud/provider published list rates plus the platform markup — never invented. For on-demand GPU/CPU the current markup is 100% (retail = 2x the cloud on-demand list rate); QPU prices follow each provider’s published pricing. The live source of truth is the pricing catalog: GET /api/v1/platform/quantum-instances/pricing-catalog returns provider prices with source URLs merged with the platform’s own rates, and the console’s Quantum Computers view renders the same data.

​
Target overview

Every workspace has the simulator and the GPU/CPU tiers in its default catalog; real QPU targets require an entitlement for the workspace — running a target you are not entitled to returns 404 no_active_entitlement_for_offer. Owners/admins can enable or disable individual targets for their workspace.

​
Simulators — free

The statevector simulator (qcos.simulator.statevector, executed by the QCOS runtime) is free for every workspace, including free-plan instances. Run the quickstart at no cost. The free CPU simulation tier (softquantus.sim.cpu) adds 10 free CPU-minutes per workspace per month. It requires a runtime bound (max_runtime_seconds); at the allowance it blocks with 403 free_cpu_limit_met and charges nothing — switch to the paid cpu.* tiers for more.

​
CPU compute

Billed per minute of runtime. Reference rates (cloud list rate x 2; the live catalog is authoritative): Run one with qcoscloud cpu, the SDK’s cpu_job() or a CPU_JOB workload. A runtime bound is required; the reservation is rate x minutes.

​
GPU compute

Billed per minute, NVIDIA GPUs, num_gpus multiplies the rate. Reference rates (cloud list rate x 2; the live catalog is authoritative): Worked example. An a100-80gb job bounded at 30 minutes reserves 30 x 17.07 ≈ 512 credits worst case. If it finishes in 12 minutes you are charged 12 x 17.07 ≈ 205 credits and the rest of the reservation is released. If GPU/CPU capacity is not currently available for your workspace, a submission fails closed with 503 gpu_provider_unavailable and no money is held.

​
Real QPUs via QCOS

Real hardware runs through the QCOS runtime. The Execution Plane recognizes devices from IonQ, Quantinuum, Pasqal, Rigetti and IBM (IBM-class also covers QuEra and IQM devices for pricing purposes).
  • Most QPUs are metered per shot at the provider’s published rate plus the platform markup. Per-shot rates vary by device and change with provider pricing — list live devices and prices with the pricing catalog, GET /api/v1/platform/qcos/backends, qcoscloud computers or qc.computers().
  • IBM-class devices are metered per second of runtime. A run is admitted only with an explicit max_runtime_seconds bound (422 runtime_bound_required otherwise) and reserves rate x bound; the run is terminated at the bound so the reservation always covers the worst case.
  • Some devices carry a provider minimum charge; when one applies, the reservation covers it so settlement can never exceed what was reserved.
Guard QPU spend per run with max_estimated_cost_usd and max_shots (rejections happen before any reservation). Access is additionally shaped by your quantum instance: free instances are simulator-only, and instance allowlists/access groups apply.

​
Cost safety

The Execution Plane is the single bank for every backend:
  • Reserve before run. A prepaid run reserves its estimated worst-case cost before anything is submitted. An insufficient balance is rejected up front (402 insufficient_credits) — nothing runs, nothing is charged.
  • Settle on completion. The actual metered cost is debited and the unused reservation released. actual_cost_usd on the job is the real charge.
  • Settlement is capped at the reservation. For prepaid runs the settled debit can never exceed what was reserved — a balance cannot go negative; the platform absorbs provider overruns.
  • Stop when unpaid. GPU/CPU jobs are metered continuously by an enforcement pass (about once a minute) and the resource is stopped and deprovisioned the moment its reservation is exhausted. Nothing runs unpaid — this is also why GPU jobs auto-stop when a wallet reaches zero cover.
  • Fail closed. An unavailable backend rejects cleanly with no money held; an unpriced target is refused rather than guessed at; reservations stranded without a provider are auto-released.
  • Free tiers never charge. They block at their caps instead of billing.

​
Per-run guardrails

Workspace-level soft/hard limits and rolling budgets are described under Billing; check your live balance any time with GET /api/v1/platform/billing/summary, qc.credits() or qcoscloud credits.