Short answer: the base audit is US$1,499. There is no single honest flat price for QPU usage because provider charges and workload requirements vary by backend, circuit or problem size, shots, repetitions, mitigation and access model. Haivian does not spend on quantum hardware until the workload and ceiling are approved.
The fee pays for scrutiny, not a positive verdict. A report can support, qualify, contradict or find the claim inconclusive. Publication is separate, optional and never guaranteed.
What the base audit covers
The initial audit freezes the claim and exact submission, inventories evidence, identifies the validation method, examines assumptions and provenance, and returns bounded findings with human accountability. Depending on scope, it can include source inspection, formal reasoning, classical simulation or review of existing provider job records.
Material scope changes, specialist work, extensive environment reconstruction and new provider-backed compute are not silently absorbed into the base fee. They require agreement.
What drives QPU cost
- Provider and backend: devices and access channels use different price models.
- Workload volume: circuit count, problem instances and parameter sweeps multiply submissions.
- Width and depth: qubit count and gate depth affect feasibility, compilation and sometimes pricing.
- Shots and repeats: statistical precision and between-run comparisons need samples.
- Compilation: transpilation, native-gate conversion and layout searches consume work and can change the executed circuit.
- Calibration and timing: a plan may require repeated execution across device states.
- Error mitigation: mitigation can multiply circuits or shots.
- Queue or priority: reserved, priority and on-demand access can have different economics; batching does not guarantee queue priority.
- Classical controls: simulators, preprocessing and baseline methods also consume compute.
Provider pricing varies
AWS Braket's official pricing page shows that on-demand QPU use can combine a per-task fee and a per-shot fee, while reserved access can be hourly. The same page illustrates how error mitigation can impose a minimum shot count. These are provider examples, not Haivian prices, and they can change.
IonQ exposes an API for estimated and actual job cost and publishes a resource estimator using variables such as qubits, gates, shots, hardware and mitigation. IonQ states that estimates are not binding. Cost evidence should therefore record both the pre-run estimate and actual provider charge where available.
D-Wave's timing documentation distinguishes QPU programming and sampling components. A workload's scientific design and the provider's billing unit must be mapped explicitly rather than guessed from a generic “quantum run” label.
The no-surprise spending control
- Freeze the exact workload and acceptance criteria.
- Identify the provider, backend and access route.
- Estimate circuits or samples, shots, repeats, mitigation and classical compute.
- Record assumptions, uncertainty and what could trigger re-estimation.
- Set a customer-approved maximum spend.
- Stop before exceeding the ceiling; do not treat an estimate as spending authority.
- Reconcile actual provider usage with the approved plan.
A provider estimate can be stale, incomplete or rejected when account limits are reached. AWS documents per-device spending limits that can prevent submission when estimated cost exceeds the remaining limit. Unknown or materially changed cost is a stop condition, not permission to continue.
When the cheapest hardware run is still waste
A low-cost QPU job can be poor value if the claim is undefined, the circuit is wrong, the baseline is weak or the output cannot be traced to the paper. Preflight proof, simulation and classical comparison reduce waste because they determine whether hardware can answer the remaining question.
Conversely, a properly scoped hardware run can be essential when the published claim is about device behaviour. The right budget follows the evidence need; it does not begin with a desire to put “run on quantum hardware” in a report.
How to request a bounded quote
Provide the exact claim, manuscript or repository version, circuit or problem description, preferred provider if any, expected backend, approximate width and depth, shots, repetitions, mitigation, available job records, required comparison and deadline constraints. Do not send secrets or private-provider credentials through a general contact form.
Haivian will first decide whether QPU execution is necessary and currently feasible. Provider-backed execution is subject to commissioned access and availability; it is not represented as an always-on included resource. Read when hardware is necessary and what a hardware audit must preserve.