Quantum computing benchmarks

There is no TOP500 for quantum computers. No independent body runs one fixed test on every machine and publishes a ranked list. What exists instead is a set of metrics, each defined by a company or a research group, most of them reported by the builder of the machine being measured. Of the 93 records here, 71 are vendor reported. This page lists them, says what each measures, and keeps the basis visible so a vendor's own figure is never mistaken for a verified one.

Longer read: There is no TOP500 for quantum: what each benchmark is worth.

Why a single ranking does not exist

The TOP500 works because every entrant runs the same program, High-Performance Linpack, on the same kind of arithmetic, and a third party publishes the result twice a year. Quantum hardware breaks each part of that. Different qubit technologies have different strengths, so no single test favours all of them fairly. A machine's useful capacity depends on its error rates and connectivity as much as its qubit count. And the test has to scale: a benchmark that a classical computer can simulate stops being informative once the machine is large, while one that cannot be simulated cannot be checked.

So the field reports several different things, and they answer different questions. Comparing them across technologies, or adding them up, is not meaningful. The tables below rank machines within one metric only, using the best figure each machine has published, and the basis column is the part to read first.

Who collects quantum benchmarks

Several efforts gather or define benchmarks. None of them is a ranking, and the ones that publish results say so themselves.

Effort	What it is	Does it rank machines?
Metriq (Unitary Foundation)	An open repository of benchmark runs with a composite score. Its March 2026 paper scores 11 devices from 5 vendors; the live dataset is dominated by IBM runs and has none from Google or neutral-atom machines.	Scores, but its authors say the result should not be read as a definitive or current ranking.
DARPA Quantum Benchmarking Initiative	A staged programme that vets whether any approach can reach utility scale by 2033. Stage B (November 2025) named eleven companies for a year of deeper evaluation.	No. DARPA describes it as not a competition between companies.
QED-C application-oriented benchmarks	A public suite of application circuits measuring fidelity against circuit width and depth. IonQ's algorithmic-qubit metric builds on it.	No. The repository publishes code, not a leaderboard.
Quantum advantage tracker (IBM, Flatiron Institute, Algorithmiq, BlueQubit)	An open tracker, launched February 2026, that compares advantage candidates against the best classical methods.	No. It tracks claims, not machines.
Other suites: SupermarQ, Q-Score, BACQ, MQT Bench	Benchmark definitions and circuit libraries from industry and academic groups.	No.

Reviews that make the same point from inside the field: a Nature Reviews Physics perspective on the "incomparable and complementary benchmarks" in use, and a review by NPL and the UK's National Quantum Computing Centre, which finds that each manufacturer can still choose which metric to report. A 2024 proposal argues for a SPEC-style standards body; none exists yet.

The only metric with a long record: quantum volume

Quantum volume was introduced by IBM researchers in 2018. Almost all of its recent record-setting reports come from one company, Quantinuum, which has published a progression to 2²³ in May 2025 and 2²⁵ (33,554,432) in September 2025, each announced by the company. The metric has known limits: IBM's own layer-fidelity paper calls it a discrete pass/fail that looks only at the best subset of a device, and its classical verification cost grows exponentially, which is part of why other vendors have moved to different measures.

Every headline advantage claim has been challenged

Google, 2019. A 53-qubit sampling experiment disputed by IBM within days, and later reproduced classically on GPU clusters.
IBM, 2023. The "utility" experiment on a 127-qubit processor was matched classically with tensor networks within two weeks. IBM's own framing is that it was not a claim of beating classical computers.
D-Wave, 2025. A quantum-supremacy claim in Science met several classical rebuttals, including a tensor-network simulation later published in Science. D-Wave disputes that its result was overturned.
Google Willow and USTC Zuchongzhi 3.0, 2024. Large random-circuit-sampling claims (Google, USTC). No published classical rebuttal was found in our research, which says the claims are untested, not that they are confirmed.

Why error correction is the number everyone watches

Because raw qubit counts and sampling claims are hard to compare, progress is increasingly framed as error correction: whether adding more physical qubits to a code makes the logical error rate fall. Google's distance-7 surface code (December 2024) is the clearest published example, with the logical error rate falling by a factor of 2.14 at each step up in code size. Logical-qubit counts from different groups are not comparable: they mix error detection with true correction, post-selection and codes of different strength.