In September 2026, researchers led by Sandia National Laboratories scored processors from Google, IBM, and Quantinuum on one benchmark. With their QUOPS benchmark, the authors gave each machine two numbers, a score for the size of the largest circuit it runs successfully and a rate for how fast it runs. Google’s Willow ran 20 million of the benchmark’s operations per second, about 66,000 times the rate of Quantinuum’s Helios-1. Helios-1 ran the larger circuits, with a score of 1,504 against Willow’s 216, the authors reported in a preprint that has not been peer reviewed.
Willow is a superconducting chip with an error-correction cycle of about 1 microsecond. Helios-1 is a trapped-ion computer with 98 physical qubits, any two of which can interact directly. Google and Quantinuum each lead their public materials with the axis on which their own machine is strongest.
Vendors headline different metrics in large part because their machines lead on different axes, and the numbers are unlikely to converge until buyers require a common test. Disputes over benchmark rules, and one word used for several measures, widen the gap. The nearest things to a common test are government evaluations run by the Defense Advanced Research Projects Agency (DARPA) and the Department of Energy (DOE). Neither agency has published a score that a buyer could set beside a vendor’s own figure.
Six headline numbers
On their own pages as of November 12, 2026, six hardware developers led with six different quantities. IBM, IonQ, QuEra, and Microsoft led in part with machines they have not yet finished. Microsoft also uses a level from its own classification scheme. Some quote logical qubits, each a unit of quantum information that an error-correcting code spreads across several physical qubits.
| Company | Headline measure | Leading figure | Source |
|---|---|---|---|
| IBM | Planned operations; programmable and physical qubits, circuits per second | A plan to link two Nighthawk r3 processors to run 10,000 quantum operations; Nighthawk r2 at 120 programmable and 458 physical qubits, 109.7 kHz | IBM hardware page |
| Verifiable advantage, error correction | Quantum Echoes, about 2 hours on Willow against an estimated 13,000 times longer on a supercomputer | Google Research | |
| Quantinuum | Gate fidelity, quantum volume | Helios, 99.921% two-qubit fidelity, and System Model H2, quantum volume of 33,554,432 | Helios page, quantum volume page |
| IonQ | Two-qubit fidelity, projected logical qubits | 99.99% fidelity, and a future system of more than 2 million physical qubits and 40,000 to 80,000 logical qubits | IonQ homepage |
| QuEra | Logical qubits at a target error rate | Libra, planned for 2028, 256 logical qubits at a logical error rate of 1 in a million | QuEra |
| Microsoft | Qubit stability, a level in its own classification, logical qubits | Majorana 2 chip; Magne, “the world’s first Level 2 quantum computer,” being built with Atom Computing, with 50 logical qubits on more than 1,200 physical qubits | Microsoft Quantum, Atom Computing |
None of the six companies publishes a conversion from its figure to any other’s. A buyer with all six figures still cannot rank the six machines on one scale.
Speed, size, and accuracy
Google added a neutral-atom program to its superconducting work in March 2026 and has described the trade between the two in its own words. In a post that month, Hartmut Neven, who leads Google Quantum AI, wrote that “superconducting processors are easier to scale in the time dimension” and “neutral atoms are easier to scale in the space dimension.” He equated time with circuit depth, the number of steps a machine can run, and space with qubit count.
Neven wrote that a superconducting cycle “takes just a microsecond.” Neutral atoms, he wrote, make up for cycle times measured in milliseconds with connectivity between any pair of atoms. He added that researchers have built neutral-atom arrays of about 10,000 qubits.
The Sandia-led authors put numbers on the same trade. They calculated that, at the error rates in their model and without error correction, physical qubits linked to four neighbors each would limit a machine to about 330 QUOPS. Full connectivity would cap it at about 10,000 “no matter how many physical qubits were available,” they wrote.
Willow, with a square lattice, scored 216, below the first ceiling. Helios-1, with full connectivity, posted 1,504, above it. The superconducting chips led on rate, with 310,000 QUOPS per second for IBM’s ibm_boston and 20 million for Willow, against 353 for Quantinuum’s H2-1 and 303 for Helios-1, as Quantum Observer reported in its article on the QUOPS benchmark.
How each company keeps score
IBM
IBM introduced a framework in July 2026 built on three quantities. They are programmable qubits, qubit operations (which IBM calls QuOps), and maximum circuits per second. IBM wrote that it expects Nighthawk r2 “to deliver 25x the circuit throughput of today’s Heron fleet.” On its hardware page, IBM lists Nighthawk r2 at 109.7 kHz in that measure, or about 110,000 circuits per second, roughly 28 times the 3.9 kHz it lists for Heron.
On the same page, under the label “Quality,” IBM lists an error per layered gate of 3.48 in 1,000 for Nighthawk r2 and 2.38 in 1,000 for Heron. By IBM’s own figures, the newer chip leads on speed and trails on that measure of quality. On its cloud console, IBM still shows CLOPS, an older IBM throughput measure, at 340,000 for the Heron processor ibm_boston, beside 3.8 kHz in the newer measure.
IBM has also changed its counts and its labels. In its April 2025 roadmap, IBM planned up to nine linked Nighthawk modules of 120 qubits each, 1,080 in all, for 2027. In November 2025, IBM said Nighthawk-based systems could have “1,000 or more connected qubits” by 2028. On November 11, 2026, IBM said it plans to link two Nighthawk r3 processors into one system of 1,040 qubits in 2027.
IBM now lists Nighthawk r2 at 120 programmable and 458 physical qubits. In July, it wrote that those programmable qubits are “supported by hundreds of additional physical-qubit resources,” such as coupler qubits. IBM also wrote that “programmable qubits and logical qubits will be used interchangeably” once logical circuits dominate. The company added that “T-gate counts are expected to replace two-qubit operations” as the measure of the hardest operations a machine can run.
Quantinuum
Quantinuum leads its Helios product page with accuracy, listing gate fidelities of 99.9975% for one qubit and 99.921% for two along with “98 fully-connected qubits” and “50 logical qubits.” On its quantum volume page, Quantinuum wrote that its System Model H2, with 56 physical qubits, “achieved a record quantum volume of 33,554,432 in September 2025.”
IBM researchers defined quantum volume in a 2019 paper as a measure of “the largest random circuit of equal width and depth that the computer successfully implements.” A machine scores higher with better fidelity and connectivity, and the time a circuit takes does not count. A register in which every qubit can interact with every other, with two-qubit fidelities above 99.9%, suits that test.
Google leads with error correction and with what it calls verifiable advantage. Its team reported in December 2024 that Willow, with 105 physical qubits, ran a surface code, a type of error-correcting code, below threshold. With each step up in code size, the logical error rate fell by a factor of 2.14. In October 2025, Google researchers reported Quantum Echoes, an experiment that took about 2 hours on Willow and that they estimated would take “13,000 times longer” on a classical supercomputer.
IonQ
IonQ set its performance targets in algorithmic qubits, or #AQ, and announced #AQ 64 on its Tempo system on September 25, 2025, “three months ahead of schedule.” IonQ wrote that “with each increment in #AQ value, the useful computational space for running quantum algorithms doubles.”
Quantinuum, another trapped-ion company, criticized the metric in March 2024. It wrote that #AQ is “much easier to pass than QV and is a poor substitute for measuring performance.” Quantinuum wrote that compiler optimization and error mitigation, a family of techniques that reduce the effect of noise on results without correcting errors, are “typically used when sharing algorithmic qubits results.” It said the benchmarks of the Quantum Economic Development Consortium (QED-C), which #AQ draws on, forbid both.
IonQ and Quantinuum build the same kind of machine and still disagreed on the scoreboard. Their dispute was over whether a benchmark should credit error mitigation and compiler optimization. That part of the divergence has no architectural cause.
IonQ has since moved its headline away from #AQ. In April 2026, it introduced a framework of 13 benchmark workloads. At the web address it used for algorithmic qubits, IonQ now presents that framework without mentioning #AQ. IonQ scores every benchmark on solution quality and gives time to solution for five of them. IonQ wrote that it evaluates all performance management layers in each benchmark, including error mitigation and circuit optimization, the practices Quantinuum objected to in 2024.
On its homepage, IonQ features what it calls a “world record two-qubit gate fidelity” of 99.99%. It describes a future system of “over 2,000,000 physical qubits” that would “translate to between 40,000 and 80,000 logical qubits.”
QuEra
QuEra leads with logical qubits at a stated error rate. On its website, QuEra plans Libra for 2028 with 256 logical qubits on more than 10,000 physical qubits, at a target logical error rate of 1 in a million. It plans a next system, for 2028 or 2029, with more than 1,000 logical qubits at a target of 1 in a billion.
Harvard and QuEra researchers reported in November 2025 that they had run up to 96 logical qubits of distance 4 at once on arrays of up to 448 atoms. Such a count is consistent with Neven’s account, in which neutral atoms scale most easily in qubits.
Microsoft
Microsoft scores machines against a classification of its own, with different metrics at each of its implementation levels. Microsoft measures Level 1, today’s noisy machines, by counting qubits and by quantum volume. Microsoft measures Level 2 by logical error rates and Level 3 by a new measure.
On its homepage, Microsoft now leads with its Majorana 2 chip, which it says “achieves unprecedented qubit stability.” Further down, it presents Magne, which it calls “the world’s first Level 2 quantum computer.” Microsoft gives no figures for either on the page. Atom Computing, which is building Magne in Copenhagen with Microsoft, described the machine in July 2025 as having 50 logical qubits and more than 1,200 physical qubits, expected to be ready for its first tasks by the start of 2027. In September 2026, The Quantum Insider reported that Magne was expected to be fully operational in spring 2027.
The Level 3 measure is for machines that do not exist yet. Microsoft describes rQOPS, reliable quantum operations per second, as “the number of logical qubits in a quantum system multiplied by the logical clock frequency of the system,” quoted with a logical error rate. In 2023, Microsoft wrote that “the first quantum supercomputer will need to deliver at least one million rQOPS,” with errors of no more than 1 in a trillion.
Microsoft has reported no rQOPS figure for its own hardware. Outside researchers have disputed data behind its 2025 topological-qubit paper.
One word, four meanings
Companies and laboratories have also given one word four meanings. In its July framework, IBM uses QuOps for the two-qubit operations it counts as a measure of quality, although its hardware page labels error per layered gate as quality. Sandia’s QUOPS stands for Quantum Universal Operation Performance System, a benchmark with a score and a rate.
Infleqtion described its 30-logical-qubit circuit as about 1,000 physical operations, or “1 KiloQuOp.” QuEra calls the class of machine it plans for 2028 “megaquop.”
The case for convergence
The Sandia-led authors make the strongest case that one yardstick can serve every machine. They wrote that QUOPS allows “fair comparison across disparate platforms,” and they ran it on hardware from three companies and on an encoded processor. Microsoft makes a related case for rQOPS once machines are fault tolerant, and IBM expects programmable and logical qubits to merge as measures of scale.
On that view, the spread is a stage the industry will pass through. Once machines correct their own errors, every vendor would report the same three numbers, a count of logical qubits, an error rate for them, and a rate of reliable operations per second.
The Sandia-led authors themselves report two figures for every machine, a score and a rate, with Willow ahead on the rate and Helios-1 ahead on the score. Vendors would still differ once machines correct their own errors, on clock speed if not on count. Microsoft’s rQOPS is the product of the two. A neutral-atom machine with millisecond cycles could lead on logical qubits and still trail a superconducting machine on clock speed by a factor of about 1,000.
Twenty of the benchmark’s 30 authors list Quantinuum among their affiliations. In its third-quarter results on November 5, Quantinuum said it had introduced QUOPS with Sandia National Laboratories and NVIDIA. Its Helios page still leads with fidelity and logical qubits. None of the other five companies cited QUOPS on the pages listed in the table.
Where the gap costs most
Planners who estimate when a quantum computer could break today’s public-key cryptography face the hardest version of the problem. For 2027, IBM plans a system of 1,040 connected qubits built from two Nighthawk processors, and for 2028 QuEra plans 256 logical qubits. For 2029, IBM has said its Starling machine will run 100 million operations on 200 logical qubits. A planner cannot compare those targets without converting connected qubits into logical ones and reconciling QuEra’s error-rate target with IBM’s operation count.
Craig Gidney of Google Quantum AI estimated in 2025 that a machine with 897,864 physical qubits, reported as fewer than 1 million, could factor a 2048-bit RSA key in about 5 days, with 1,409 logical qubits in use at its peak. He assumed a gate error of 0.1% and a surface-code cycle of 1 microsecond. He also gave the control system 10 microseconds to react.
The Sandia-led authors converted that estimate into a target of about 250 million QUOPS, run at about 5,700 QUOPS per second to finish in 5 days. They wrote that today’s physical-qubit processors have, or nearly have, the necessary rate, and that current scores fall short by roughly five orders of magnitude. Shor’s factoring algorithm needs a machine that is both large and fast enough, so a planner has to track logical-qubit counts and clock speed together.
The nearest common yardstick
DARPA’s Quantum Benchmarking Initiative comes closest to a common yardstick. In Stage C, its final phase, a government team checks that a company’s proposed utility-scale machine can be “constructed as designed and operated as intended,” DARPA says. As of October 8, 2026, DARPA listed six companies in Stage C, as Quantum Observer reported in its article on the Stage C selections. Google Quantum AI, which joined the first stage in September 2025 according to DARPA’s program page, is not among them.
Atom Computing uses neutral atoms, Diraq silicon spins, IBM superconducting circuits, IonQ trapped ions, Microsoft topological qubits, and PsiQuantum photons. DARPA evaluates each approach separately and shares its results with other U.S. government agencies. It does not publish a common number that buyers could cite.
The Energy Department has written the divergence into its rules. In the Quantum Genesis Q Competition, DOE intends to use national-laboratory staff, and possibly other neutral third parties, to verify performance. For architectures “not naturally measurable in total logical qubits,” DOE wrote that it will “identify equivalent technical metrics during negotiation” with each company, as Quantum Observer reported in its article on the competition.
Standards bodies and industry groups in Europe and the United States have started on the same problem. CEN and CENELEC’s joint quantum committee, JTC 22, is preparing a technical report on performance benchmarks of quantum computing applications, a project that Germany’s DIN lists as started in March 2024. An IEEE working group, P7131, has a project on quantum computing performance metrics but has published no standard. QED-C published an open-source application benchmark suite in 2021.
Until a common test exists, buyers can ask every vendor for the same four items. They are the circuit family behind a result, the compilation and error mitigation applied, the fidelity across the whole register as well as the best pair, and how long, in wall-clock time, the machine ran to produce the result.
What would change this
This reading would be overturned if the first company to reach fault tolerance drew competitors toward its architecture and its metrics. It would also be overturned if DOE or DARPA published evaluation results in one format for buyers to cite in place of vendors’ figures.
The next tests are scheduled. Quantinuum expects to deliver Apollo, its universal fault-tolerant machine, by 2029, the year IBM has set for Starling.
If the companies report those machines in the units they use today, buyers will still be comparing different quantities.
