Verified quantum advantage: IBM certifies a 70-qubit result

What IBM actually claimed

For most of the past decade, claims of quantum advantage had a hidden weakness. A quantum processor would produce output that no classical machine could reproduce in reasonable time. Because nothing else could reproduce it, nobody could easily confirm that the output was correct. IBM and the University of Chicago now say they have closed much of that gap. Their claim is that a quantum computation went beyond leading classical simulation methods and came with statistical evidence that the answer was right.1220

The paper is titled "Sampling hard circuits with verifiably high fidelity." It appeared on arXiv two days before IBM's July 30, 2026 announcement. A later wave of coverage misdated the story to late August, when ScienceDaily republished IBM's press release.19 IBM Research director Jay Gambetta framed the result in bold terms, saying the field is "now firmly in the quantum advantage era." He said the experiment set a statistically confident lower bound on how faithfully the computation ran.1220

The circuits and results have been posted to IBM's public Quantum Advantage Tracker so outsiders can benchmark and attack them.20 In this field, that kind of openness matters a great deal.

How the verification works

The key idea uses a known asymmetry in quantum computing. Circuits built only from Clifford gates are easy for classical computers to simulate, no matter how many qubits they use.1617 The IBM–Chicago team first built a circuit entirely from Clifford gates and wrapped it in an error-detection scheme they call a "spacetime code." At that stage, the circuit's fidelity can be checked directly against a classical simulation.1719

Next, the team added 468 non-Clifford "T gates" to a 70-qubit circuit. The T gates are what make the circuit hard to simulate. The authors say 468 is more than twice the level at which classical simulation stops being feasible.1117 Error syndrome statistics looked the same before and after the T gates were added. From that, the researchers bounded how much fidelity the added gates could have cost.19

The measured fidelity of the Clifford reference circuit was 0.32. Subtracting the worst-case penalty gave a certified floor of 0.284 at 95% confidence.1419 The team ran 2,415 logical two-qubit operations on a Heron-class processor. The whole run took about 15 minutes, and effective error rates came out roughly ten times lower than the underlying physical rates.1317

The 0.284 figure looks modest, but it should be read correctly. It is a conservative lower bound, not a point measurement. With readout error mitigation, the authors estimate the true state fidelity at about 0.57.19 IEEE Spectrum noted that no previous approach had offered a 95% confidence guarantee of this kind.11 In my view, that guarantee is the real news. The speed of the computation is secondary.

Where the coverage diverges

Coverage of this experiment is not unanimously positive, and the details differ depending on who is describing it.

The first disagreement is about what was built. IBM's release describes one of the largest error-correction demonstrations to date, running 70 logical qubits. A careful reading of the paper shows something narrower: 70 data qubits plus 27 ancilla qubits, 97 physical qubits in total, protected by error detection with post-selection, not full error correction.1920 That difference matters. Error detection throws away bad runs instead of fixing them. Roughly one run in 1,700 survived the checks, and the full process needed about 860 times the normal number of shots.1719

The second disagreement is about how far beyond classical reach the result really is. On August 13, researchers at Singapore University of Technology and Design, working with an NVIDIA engineer, computed exact probabilities for all 2,051 output samples IBM had published. It took 37.3 minutes on 256 H100 GPUs.19 Their method exploited the circuit's one-dimensional chain layout and its exclusive use of CZ gates.19 The surprising part is that this classical attack ended up supporting IBM's accuracy claim. The Singapore team's fidelity estimate of 0.35 is consistent with IBM's certified floor.1619

IBM then rebuilt the experiment around a harder circuit. The researchers who found the shortcut are reportedly co-authors of the revised version. That paper is also a preprint, and the claim is still being tested.16

My reading is that the speed claim has softened, but the verification claim has held up. A 15-minute quantum run against a roughly 37-minute GPU calculation is not the dramatic gap the phrase "beyond the practical reach of classical computers" suggests. Still, the certificate itself survived the first serious attack. Commentators who call this healthy are right: a falsifiable claim invited the strongest classical response, and the claim was refined instead of abandoned.16

The other two experiments

The UChicago paper was released alongside two partner studies, and these deserve more caution. Qedma used IBM's Heron to simulate oscillating quantum dynamics. Up to 35 qubits, quantum and classical methods agreed. At 51 qubits the classical methods broke down, and at 74 qubits they could not handle the problem at all.1117 Qedma's strongest check was running the same simulation on Quantinuum's trapped-ion machines and seeing the same behavior. Those machines have completely different noise profiles from IBM's superconducting chips.1117

Algorithmiq ran a 56-qubit circuit and found that competing classical methods contradicted each other in the hardest regime. The team then cross-checked the quantum results with shortened circuits, deliberately injected noise, and a second processor.17 Chemistry and Engineering News pointed out a weakness here. Both quantum processors share the same architecture, and the simulated material was designed for IBM's hardware rather than taken from a real substance. That means verification remains largely internal.14

ETH Zurich's Dominik Hangleiter noted that neither partner paper explicitly claims quantum advantage, and he called that restraint appropriate.17 He went further, saying it is unlikely anyone will ever prove quantum advantage definitively.14 All three papers were still preprints awaiting peer review at announcement time.17

The competitive picture for quantum companies

IBM's push is best understood as positioning in an increasingly crowded race over who gets to define trustworthy quantum results.

Google's earlier Willow random-circuit-sampling result was widely criticized as an artificial benchmark. In October 2025 Google followed it with Quantum Echoes, which it described as the first verifiable quantum advantage on hardware, running 13,000 times faster than its chosen classical comparison.110 A 2026 preprint, with several authors affiliated with Google, argues that belief-propagation tensor networks cannot feasibly simulate that experiment.16 Even so, Google's own framework acknowledges that no end-to-end quantum application has yet shown a conclusive advantage on a problem of real-world consequence.10

Quantinuum took a different approach to the same verification problem. Its team, led by Marcello Benedetti and Harry Buhrman, designed a game with a mathematically proven ceiling on how well any classical strategy can perform. Its H2 trapped-ion machines beat that ceiling using up to 55 qubits, and the gap widened exponentially as the problem grew.18 Unlike IBM's certificate, this test does not depend on assumptions about computational complexity.18 That makes it arguably cleaner in principle, though it is further from practical computation.

Other vendors are working on the error-correction foundations. IonQ announced on September 22 a real-time error decoder that runs on a single standard CPU. In benchmarks simulating up to 408 logical qubits and more than 31.5 million operations, it added as little as 0.02% delay.8 IonQ shares rose after the announcement.5 Independent analysts noted that these were simulated benchmark circuits, not a physical machine with 408 logical qubits.4 Infleqtion reported 30 entangled logical qubits on its neutral-atom Sqale platform.2 Separately, a BlueQubit-led team showed that random-circuit sampling beyond classical reach can now run on IBM's commercially available Nighthawk r2 processor. It generated a million samples in 19 seconds, a task estimated to take a supercomputer about 110 years.7

Why it matters

The broader shift is that quantum advantage is turning from a single headline moment into an ongoing process of evidence, challenge, and revision. Freie Universität Berlin's Jens Eisert described verification beyond classical reach as building trust through a portfolio of complementary methods rather than passing one test.17

For companies, this changes what counts as a credible claim. Raw speedups on random circuits are no longer enough, and announcements now need a way for outsiders to check the result. IBM has the strongest current example of certified correctness on a hard sampling task. Its marketing language about "logical qubits" and "error correction" goes beyond what the paper supports, and its speed gap shrank quickly once classical researchers took it on.19

None of these experiments solves a commercially valuable problem yet, and experts across the coverage agree that commercial usefulness is still years away.14 What has changed is the standard for what a believable advantage claim must include.