You Can Compute It — But Can You Verify It? Quantum Computing's 2026 Verification Problem

Through 2026, IBM, Google, and Microsoft/Quantinuum have each announced "progress" measured on entirely different scales. But can these three kinds of progress really be lined up on the same ruler? A look across this year's announcements suggests the real barrier to practical quantum computing isn't computation itself — it's the harder question of how you verify that a result is actually correct.

"Verifiable" quantum advantage — IBM's 70 logical qubits

On July 30, 2026, IBM and the University of Chicago announced results from a preprint posted to arXiv, "Sampling hard circuits with verifiably high fidelity." Using 70 logical qubits (error-corrected logical qubits, built from a larger number of physical qubits), the team sampled from structured circuits rather than random circuit sampling (RCS), the benchmark most past quantum-advantage claims have relied on. The quantum computation took about 15 minutes. IBM says many leading classical simulation approaches would face prohibitive runtimes, though it hasn't given a specific comparison runtime. In the announcement, University of Chicago's Bill Fefferman said verification "remains one of the biggest challenges in firmly establishing experimental quantum advantage."

That word — verification — carries weight because Google's 2019 Sycamore quantum-advantage claim was repeatedly chipped away at by subsequent classical algorithm improvements that turned out to simulate it in practical time after all. Whether IBM's new claim actually holds up will depend on independent replication and closer scrutiny of the verification method itself.

Letting AI take the wheel — Google's reinforcement-learning error correction

In July 2026, a Google team published in Nature a method that uses reinforcement learning (RL) to keep tuning the controls of error correction while it runs. Until now, as a device drifts over time, error correction has had to be halted for recalibration. Here, more than 1,000 control parameters are adjusted continuously without stopping the computation. In experiments with deliberately injected drift, the stability of the logical error rate improved 3.5-fold, and RL fine-tuning suppressed the logical error rate itself by a further 20% or so. Numerical simulations confirmed the approach works at a scale of tens of thousands of control parameters. The paper frames it as a step toward algorithms that need to run continuously for days or even months.

Fewer qubits, sharper precision — Microsoft and Quantinuum's error-correcting codes

In June 2026, Microsoft and trapped-ion specialist Quantinuum published in Nature results from two custom error-correcting codes, the "carbon code" and the "tesseract code" (the experiments were first reported in 2024). The carbon code protects 2 logical qubits using 12 physical qubits. According to press coverage, it cut the error rate on a Bell-state experiment from 0.8% unprotected down to 0.17%; when runs showing signs of an error were also discarded, more than 15,000 accepted trials produced zero observed errors, for an estimated error rate of 0.001% — roughly an 800-fold improvement. Run through up to 10 correction cycles, it reached a per-cycle error rate of 0.006% against physical baselines of 0.37–0.59%. The tesseract code protects 4 logical qubits with 16 physical qubits; in an experiment handling up to 12 logical qubits at once, it cut errors 22-fold compared with the unencoded circuit.

Three kinds of "progress" that can't be compared

Line these three results up and it becomes clear they're measuring fundamentally different things. IBM is tackling whether a claimed advantage can be verified by outside parties. Google is tackling whether error correction can run autonomously, for long stretches, without human intervention or downtime. Microsoft and Quantinuum are tackling how low a logical error rate can go with a modest number of physical qubits. Each is meaningful progress, but none of them alone adds up to "quantum computing has become practical." Verifiable advantage, autonomous continuous operation, and low error rates that hold at scale — all three need to come together, on the same architecture, before the field gets genuinely close to useful.

So when does this actually become useful?

IBM has set a goal of building the world's first large-scale fault-tolerant quantum computer by 2029. Looking at where things stand in 2026, though, each company's progress is still piling up separately, largely along the lines of whatever its own research group happens to be best at. The more useful lens on "how far away is practical quantum computing" may not be which hardware modality — superconducting, trapped-ion, neutral atom — eventually wins, but how far these three separate threads (verification, autonomous operation, low error rates) can be woven together on a single machine. As of September 2026, that weaving is still very much in progress.