The work around
the quantum circuit.

A faster quantum routine becomes a practical gain when the work surrounding it can deliver a useful, sufficiently reliable answer at a competitive cost. That surrounding work belongs inside the comparison.

My mini-review, From Quantum Speedup to Practical Advantage, published on 14 October 2026, separates four questions: what can be computed, how efficiently an algorithm operates, what a reliable machine must do, and what the completed result is worth. Moving between those questions requires evidence about the system that connects them.

Specify the answer being bought

A quantum state has a mathematical description whose size can grow exponentially with its number of qubits. Measurement gives access to particular outcomes. The useful engineering question is which information an application needs, and how the circuit makes that information recoverable.

The mini-review treats Grover's search algorithm and Shor's factoring algorithm as genuine algorithmic achievements. It also explains why the size of a state space does not supply a freely readable collection of all possible answers. A proposed application must identify its output: a sample, an estimate of an observable, a candidate solution or some other specified result.

That choice determines part of the workload. Estimating an average to a stated accuracy may require repeated measurements. A candidate solution may need a separate check. A useful comparison carries those requirements into both the quantum and classical routes.

Put the surrounding work on the clock

Consider a deliberately hypothetical calculation with matching inputs, output requirements and reliability. Both routes spend 20 seconds on common preparation and checking. The classical route then spends 80 seconds on its computational routine, for a total of 100 seconds.

Suppose a quantum routine takes four seconds, but its route also needs 90 seconds of additional state preparation, repeated sampling and readout. Assume, for this simple example, that these stages run sequentially and that none is already included in the common 20 seconds or the four-second routine. The total becomes 114 seconds. The computational routine is twenty times faster; the complete route takes fourteen per cent longer.

These are invented timings, not measurements of a processor or an estimate for a particular algorithm. Their purpose is to make the accounting boundary visible. If the additional work were smaller, reusable or overlapped with other stages, the conclusion could change. Those possibilities belong in the measured workflow.

Repeated use creates a further question. A setup cost incurred once for a batch has a different effect from a cost incurred for every answer. A benchmark should report the batch size and the reuse it assumes. Otherwise a one-off demonstration and a sustained service can appear comparable while carrying different costs.

Keep the comparison capable of improving

The classical alternative is part of the research problem. In a 2024 study, Tomislav Begušić, Johnnie Gray and Garnet Kin-Lic Chan examined observables from a particular 127-qubit kicked-Ising experiment. Their approximate classical methods reproduced the studied observables faster than the quantum experiment and could converge beyond its reported experimental accuracy.

The scope of that result matters. It concerns specified observables and circuits. It does not establish that every quantum computation can be replaced efficiently. It does show why exceeding a brute-force classical method is an incomplete comparison when a more suitable classical method can answer the actual question.

A benchmark therefore needs a date, a workload, an accuracy target and an account of the methods considered. Improvements on either side can change the result. Retaining the original inputs and evaluation procedure makes that change examinable.

Let engineering progress change the assessment

The mini-review also rejects a categorical impossibility claim about scalable fault tolerance. It recognises Google Quantum AI and Collaborators' below-threshold quantum-memory experiments as substantive progress, while distinguishing a demonstrated memory from the complete operations required by an application-scale computer. An assessment has to leave room for evidence that changes its premises.

For practical advantage, I propose starting with a named task and an explicit acceptance test. Record preparation, compilation, calibration, correction or mitigation, sampling, decoding and verification where they apply. Identify which costs are shared, repeated or amortised. Compare elapsed time, energy and financial cost separately before making an economic judgment; they measure different consequences.

Verification deserves its own design. If an exact classical calculation is too expensive, the evaluation still needs a justified way to assess the required output. Confidence should come from the specified test and its assumptions, rather than from the difficulty of reproducing the computation.

The resulting research programme is constructive. It asks where an algorithmic improvement survives implementation, where another stage limits it, and which engineering change would move the boundary. That is how a promising circuit becomes a credible claim about a useful system.