Five per million, doing what

Start with what was actually measured, because the press release does not volunteer it in the headline sentence. "Five errors in a million operations" sounds like a computer computing, but the operations in that framing are syndrome extraction cycles — error-checking rounds — not logical gates. The paper describes a memory experiment: a logical qubit is prepared, put through twenty rounds of error checking, and then read out. The record rate, a logical memory error rate of 5 (+4/−2) ×10⁻⁶ per syndrome extraction cycle at code distance 7, is the chance of failure per error-checking round in that setup, at the largest code size the team ran. Storing quantum information without it decaying is a real and necessary job: error-corrected memory is a component of any fault-tolerant computer. But the headline figure belongs to the memory experiment, not to executing logical gates.

The harder test sits in the same abstract. When the team ran randomised circuits built from two-qubit logical Clifford gates, the encoded operations people would actually compose calculations from, the error rate at distance 5 was 7 (+3/−3) ×10⁻⁵ per gate. That is about 14 times the numerical memory error rate, but the ratio is illustrative rather than like-for-like: the two numbers use different error definitions and were measured at different code distances. New Scientist, reporting the same paper, describes the gate rate more loosely as roughly ten times higher than the record figure, about seven errors per 100,000 uses. Neither number is hidden. But a reader who only meets the release will carry away the memory number labelled as operations, and that label does work the data does not.

The number that earns the announcement

Strip the headline back and a stronger claim remains, one that does not need the unit blurred. As the code distance rose from 3 to 5 to 7, logical memory errors fell by a factor of about twelve at each step. That is below-threshold error suppression in practice: adding protection reduces errors faster than the extra hardware introduces them. A one-off low number could be a tuning accident. A suppression factor of about twelve, repeated across the tested distances, is the scaling behaviour the whole approach depends on. The gate circuits showed the same pattern at a factor of about nine. None of that, by itself, establishes fault tolerance at arbitrary scale or for arbitrary circuits; it shows the tested experiments behaving the way the theory requires.

Two engineering details support taking the scaling seriously. The team attacked leakage, where a qubit slips out of its working states and spreads errors to its neighbours, at both hardware level, with a laser repump, and circuit level; the hardware scheme alone improved the logical error rate fourfold at distance 3 and eightfold at distance 5. And the decoding was not all done offline at leisure: the paper reports neural-network decoders, a convolutional network and a transformer, working in real time about ten times faster than the hardware produces error data.

What the record cost, in qubits

Error rates this low are bought with redundancy, and the exchange rate is the part of the story that decides timelines. Two counts describe the same experiments, and they count different things. The paper counts 73 total physical qubits for a single distance-7 memory experiment; New Scientist counts 37 data qubits at that distance. Helios-1, the whole machine, has 98 physical qubits, per New Scientist, so the distance-7 experiment uses 73 of them under the paper's counting convention. Thomas Wong of Creighton University, quoted by New Scientist, puts the requirement for known useful applications at on the order of 100 logical qubits — an estimate of the destination, not a demonstrated capability, and the resource accounting of one memory experiment is not the full overhead of a production workload either.

The company's own comparison makes the trade visible from the other side. Quantinuum says its September result on the same hardware, using its Helix code, needed 10 physical qubits per logical qubit and got a higher error rate; the company puts the new rate at 10 times lower than that result, and 100 times lower than the best published result on other hardware, which it attributes to Google Quantum AI on a superconducting system. Those are the company's comparisons, their validity depends on the underlying experimental conditions, and they are not a universal ranking. Color code buys accuracy with qubits. Helix buys qubits with accuracy. Neither is free.

Quantinuum's answer is its next machines: Hayes told New Scientist the follow-up system, Sol, will be nearly twice Helios-1's size, and the company frames today's accuracy as within the target range for Apollo, its planned fault-tolerant computer, which it targets for 2029. That date is a company target, not a forecast, stated in a press release whose own forward-looking section warns that results achieved in experimental settings may not replicate at larger scale, in logical computation, or in commercial use. The company wrote the caveat itself. It is worth believing both halves of its document: the measurement, and the warning attached to it.

How to hold the claim

The outside reaction is genuinely positive, and specific. Blume-Kohout called it real, serious performance gains, and suggested a scaled version might reach useful reliability at under a million physical qubits — a projection about a future machine, not a result demonstrated on Helios. Steven Flammia at Virginia Tech called it an impressive technical advance, while noting the open question is whether rates stay this low under more complex operations. Quantinuum published the result alongside a research paper of its own; no peer-reviewed journal version was identified in the sources checked for this article, and expert comment, however informed, is assessment rather than independent replication of the experiment.

That is the correct shape of the conclusion: below-threshold error suppression that held across three tested code sizes, a gate rate whose numerical value sits an order of magnitude behind the headline on a different measurement at a different distance, and a qubit budget that remains the binding constraint. Five per million is true, as a logical memory error rate per syndrome extraction cycle. Quantum computing gets useful when comparable rates hold for many logical qubits, computing.

Sources

  1. [1] Quantinuum press release, "Quantinuum Breaks Error-Correction World Record by 10x, Validating Its Leading Path to Fault-Tolerant Quantum Computing by 2029", 19 October 2026, 08:00 ET, via CNW/Newswire — the announcement; the 10x and 100x comparisons and the Apollo 2029 target are the company's own; its forward-looking section warns results may not replicate at larger scale or in commercial useRead source
  2. [2] Lavasani et al. (Quantinuum), "Quantum error correction well below the color code threshold", research paper linked from the release — memory rate 5 (+4/−2) ×10⁻⁶ per syndrome extraction cycle at distance 7; logical Clifford rate 7 (+3/−3) ×10⁻⁵ per gate at distance 5; suppression factors Λ ≈ 12 (memory) and ≈ 9 (gates); 13 / 37 / 73 total qubits per memory experiment at distances 3 / 5 / 7; leakage-repump gains; neural decodersRead source
  3. [3] New Scientist, "Error-correction record pushes quantum computers closer to usefulness", 19 October 2026 — independent report with outside experts (Flammia, Wong); counts 7 / 19 / 37 data qubits; Helios-1's 98 physical qubits; gate rate expressed as about seven errors per 100,000 uses; the Sol follow-upRead source
  4. [4] HPCwire, "Quantinuum Reports 99.9995% Logical Fidelity on Helios Quantum Computer", 19 October 2026 — wire republication of the release, corroborating textRead source