What does it cost to run a problem on a real quantum computer?

Status as of: July 2026. Prices are list prices pulled on the dates in Sources; they change often. Verify against the provider before you budget.

The honest answer has two halves. One job is cheap. On Amazon Braket every quantum task carries a flat $0.30 fee plus a per-shot rate, and the per-shot rate spans a wide range by hardware: $0.000425 on Rigetti's superconducting Cepheus up to $0.08 on IonQ's trapped-ion Forte — a 188× gap for a single shot. A real problem is not one job. Variational algorithms (QAOA, VQE) run a classical optimizer that submits hundreds of circuits, each sampled thousands of times. Multiply it out and the same computation costs about $73 on the cheapest device and about $8,030 on the most expensive — roughly 110× apart — before you add error mitigation, which multiplies the shot count again. The price you should track is cost-per-useful-answer, and no provider quotes that.

What does a single quantum job cost?

Amazon Braket's model is the clearest to read because it is itemized: a per-task fee plus a per-shot fee. Every device below carries the same $0.30 per-task fee; the difference is entirely in the per-shot rate.

Source: Amazon Braket pricing page, retrieved 2026-07-27. Per-task fee is $0.30 on all six devices. Reservation mode bills a flat hourly rate instead of per-task/per-shot.

At these rates a one-off 10,000-shot job is trivially cheap on superconducting hardware and not so cheap on trapped ion: $4.55 on Rigetti Cepheus, about $14.80 on IQM Garnet, $100.30 on QuEra Aquila, and $800.30 on IonQ Forte (10,000 × per-shot + $0.30). That is the number people quote when they say quantum computing "costs a few dollars." It is true, and it is misleading, because one 10,000-shot job is not a workload.

What does a real problem actually cost?

A variational run is a loop: propose parameters → run a circuit → measure many shots → estimate a cost → update parameters → repeat. Take a deliberately modest configuration — 100 optimizer iterations, one circuit evaluation each, 1,000 shots per evaluation = 100 tasks and 100,000 shots. This is an illustrative calculation from the list prices above, not a measured run; real loops are often larger.

The arithmetic is 100 × $0.30 + 100,000 × (per-shot):

Same optimizer, same shot budget, same problem: about 110× between the cheapest and most expensive device. The choice of hardware, not the algorithm, dominates the bill. And this is the floor — production variational runs frequently use more iterations and more shots, and the cost scales linearly with both.

Why per-shot price is not cost-per-answer

A cheap shot is not a cheap answer. Three reasons the sticker price understates the real cost:

Error mitigation multiplies shots. Techniques like zero-noise extrapolation or probabilistic error cancellation trade extra sampling for lower bias — often several-fold more shots for the same estimate. A device with a low per-shot rate but a high mitigation overhead can end up costing more per trustworthy number than a "pricier" one.

Not everyone bills by the shot. IBM bills wall-clock QPU time: about $96/minute on Pay-As-You-Go and $72/minute on the Flex plan (minimum ~$30,000), with a free Open plan capped at 10 minutes per 28 days. Under a time-based model your cost depends on circuit depth, queue behavior and runtime — not shot count — so it is not directly comparable to Braket's per-shot table.

Some pricing is opaque by design. Quantinuum meters in Hardware Quantum Credits, HQC = 5 + (C/5000) × [N₁q + 10·N₂q + 5·N_M], where two-qubit gates are weighted 10× — but the dollar value of an HQC is not public, and enterprise access is quoted (Azure lists Quantinuum H2 subscriptions in the six figures per month). "How much does the problem cost" has no list-price answer for these systems.

This is why Rosetta's position on cost is the same as our position on advantage: the honest unit is measured, not quoted. Our 48 sealed runs to date executed on classical simulators and classical linear algebra, where the marginal cost of a shot is near zero — which is exactly why the real-hardware cost question deserves a straight answer rather than a hero-number.

What we know / What we don't know

What we know (sourced, dated). Braket's itemized per-task and per-shot list prices as of 2026-07-27. A ~110× spread across public QPUs for one identical variational workload. IBM's per-minute Pay-As-You-Go and Flex rates. The shape of Quantinuum's HQC formula. That reservation mode exists and flips the math once a device is busy enough.

What we don't know. The dollar value of a Quantinuum HQC, or negotiated enterprise rates for any provider — list price is a ceiling, not what large buyers pay. Cost-per-useful-answer: none of these numbers account for how many shots a given device needs, with mitigation, to reach a target accuracy on a given problem. The reservation-vs-on-demand crossover for a specific workload. How long any of these prices hold — they move, sometimes monthly. And the costs that never appear on the QPU invoice: classical pre/post-processing, compilation, orchestration, and engineer time, which for most pilots dwarf the quantum bill.

Rosetta Q publishes verdicts with raw, reproducible data. This is educational content, not a product claim. Prices are list prices captured on the dates cited and are not investment or procurement advice.