Quantinuum Error Correction Record Hits 99.9995% on Helios

A record set twice in two months

On October 19, Quantinuum said its Helios trapped-ion computer had reached a logical fidelity of 99.9995%. That works out to about five errors per million operations, and the company called it a world record in quantum error correction.1 The company says the error rate is ten times lower than the record it set in September 2026. It also says the result is a hundredfold better than the best published figure on any other kind of hardware, which came from Google Quantum AI's superconducting system.1 The press release cited Google's published Nature paper as that benchmark.9

The two results ran on the same machine with different codes. In September, Quantinuum used its in-house Helix code. This time it used a standard "color code."1 The trade-off between them is clear. Helix needed 10 physical qubits per logical qubit, and the color code needed up to 37 but produced fewer logical errors.2 Quantinuum CEO Rajeeb Hazra said the result proves the company has the architecture and trapped-ion technology needed for universal fault tolerance. He also said it confirms the company's path to Apollo, the fault-tolerant machine it plans to deliver by 2029.1

The reporting agrees on the basic numbers, and most outlets repeated the company's framing almost word for word.378 The more useful coverage looked at what the numbers measure. Once you read it closely, my view is that the headline fidelity figure is the least important part of this result.

The number that matters is 12

Quantinuum also showed that logical errors fell exponentially as the "code distance" went up. Each step up cut the remaining errors by a factor of 12.1 Code distance is roughly the size of the protective code wrapped around a logical qubit. A bigger code means more physical qubits and, if the hardware is good enough, fewer errors. The factor that measures how fast errors drop with each step is called Λ (lambda).

A detailed analysis by PostQuantum explains why Λ matters. The team encoded one logical qubit in color codes of distance 3, 5 and 7, which used 13, 37 and 73 physical qubits once measurement qubits are counted. The logical error per round fell from 5.5 × 10⁻⁴ to 3.8 × 10⁻⁵ to 5 × 10⁻⁶.14 For comparison, the same analysis lists Google's 2024 surface-code result on Willow at a Λ of 2.14, Google's 2025 color-code experiment at 1.56, and a USTC Zuchongzhi result at 1.40.14 Quantinuum's own blog stresses the gap with the most comparable superconducting color-code experiment, 1.56 versus 12. It argues that this lets the company reach "utility scale" error rates with smaller codes and therefore far fewer physical qubits.20

This matters more than any single fidelity figure because Λ sets how many physical qubits each logical qubit needs. That count drives a machine's cost and delivery date.14 If errors keep falling by a factor of 12 per step, a distance-21 color code would reach one error in a trillion operations. In this layout, that code needs 661 physical qubits per logical qubit. With a Λ near 2, the same target would need distance 53 and more than six times as many qubits.14

An outside expert backed the significance. Robin Blume-Kohout, co-director of Sandia National Laboratories' Quantum Performance Lab, was not involved in the study. He called it a milestone researchers have pursued for years and said the logical qubit is nearly 100 times better than the physical qubits it is built from.1 He said a scaled-up version could reach useful fault-tolerant reliability with fewer than a million physical qubits, and that the findings largely remove one of the field's main technical risks.3 Steven Flammia of Virginia Tech, also independent, called it an impressive technical advance.18

Where the headline outruns the data

The closer readings disagree with the press release in a few places, and the details matter.

It is a memory result, not a computation result. The 99.9995% figure is fidelity per round of error correction. The team prepared a logical state, ran 20 rounds of checks, and measured it, with no computation in between.14 Quantinuum's blog says so directly, calling it "logical memory fidelity."20 The press release just says "logical fidelity," and several outlets turned that into "five errors in a million operations" without saying what kind of operation.717 The figure that reflects computation is the error per logical two-qubit gate. That was 7 × 10⁻⁵ at distance 5, about 14 times better than the same gates on bare physical qubits.14 That is a strong result, but it is well short of six nines.

The "tenfold" gain is rounded up. Going from September's Helix result of 4.6 × 10⁻⁵ per round to 5 × 10⁻⁶ is about a 9.2-fold improvement. Logical gate errors improved about fourfold.14

The distance-7 point rests on very little data. The fitted Λ of 12 covers three distances. The step from distance 5 to 7 was only about 7.6, and that ratio depends on just three observed failures at distance 7.14 With only three code sizes measured, nobody can yet say whether the suppression rate holds as codes get larger.14 Quantinuum's forward-looking statements also caution that the results came from experiments and may not hold at larger scale or in commercial use.1

The qubit-count comparisons leave things out. Quantinuum estimates that 100 logical qubits would take about 30,000 physical qubits on its approach, against about 900,000 for a superconducting color code. That estimate counts only data qubits, so both absolute numbers are roughly half the real totals. The ratio between them still holds at about 27 to 1.14 Neither estimate counts the extra machinery needed for a full fault-tolerant computer, such as magic-state factories.14

Some coverage also got the physics wrong. One report described trapped-ion machines as using "ytterbium or barium" ions in general, and paraphrased Hazra as saying fault tolerance would not need millions of qubits. That claim is not in the published quote.17 Helios uses barium ions, while the older H2 used ytterbium.13 None of this undercuts the result, but it shows how quickly a careful technical claim turns into a broad one.

How the hardware got there

Part of the improvement came from better control of the physical qubits, not just the code. Quantinuum's David Hayes told New Scientist that when a qubit's energy drifts in a harmful way, it now gets hit with a corrective laser pulse before the color code does its work.18 The paper describes a "repump" step that cut the number of qubits leaking out of their working states 6.5-fold without adding time to the circuits. It lowered logical error rates fourfold at distance 3 and eightfold at distance 5.14

That sits on top of the Helios hardware Quantinuum launched in November 2025: 98 fully connected qubits with 99.9975% single-qubit and 99.921% two-qubit gate fidelity.12 Those fidelity figures have since been confirmed in a peer-reviewed Nature paper written with Sandia.13 Quantinuum's own line is that its world-record logical fidelities come from its world-leading physical fidelities.20

The roadmap: Helios, Sol, Apollo

Quantinuum says its accuracy is now "within the target range" for Apollo.2 PostQuantum's numbers make that plausible. Quantinuum and SoftBank had put Apollo's logical error target at about 10⁻⁷ at code distances of 7 to 9. If Λ stays at 12 on current hardware, Helios's memory error would fall into that range between distance 9 and 11, before any of the fidelity gains planned for Apollo.14 The analysis notes that the company did not say whether that target is per gate or per round, so this is still a rough comparison.14

The next step is Sol in 2027. Quantinuum has said Sol is designed to deliver 100 logical qubits from 192 physical ones at a logical error rate of 10⁻⁵.14 Tracking sites list Sol at up to 100 logical qubits.19 That target depends on Quantinuum's high-efficiency codes, not the color code used for this record. Helios has already run computations on 48 error-corrected logical qubits using concatenated "iceberg" codes.14 Independent reviewers note that those claims are narrower than they sound. Some of the larger logical-qubit counts depend on error detection with discarded runs, not full correction.13 Public trackers don't even agree on the corrected count, with some listing 48 and others 50.1619

The challenge for Quantinuum is to combine the two strands of its work. One code wins on accuracy and the other wins on efficiency. Apollo needs both at once and at scale. The company's architecture is designed for this: it can adjust error checks during a computation and link different codes together.8 That combination has not yet been shown on hardware.

The competitive picture

Quantinuum isn't the only company claiming leadership in error correction, and the claims depend on what you measure. By logical-qubit count, QuEra leads with 96 logical qubits from 448 neutral atoms, published in January 2026. Quantinuum holds the best encoding ratio.16 IonQ, Quantinuum's main trapped-ion rival, has claimed the two-qubit gate fidelity record at 99.9923%.19 Google was first to show that errors shrink as codes grow, in late 2024.19 The leaderboard ranks Infleqtion and Atom Computing with Microsoft as well, and several of those companies are targeting 100 or more logical qubits.16

On the specific question of how fast errors drop as codes grow, Quantinuum's result now leads clearly. Even a skeptical reading found nothing in the per-second comparison with Willow to dispute the company's claim of the most robust quantum memory ever recorded.14

Investors were more cautious. One market summary showed Quantinuum shares down about 7.6% at the time it was published. It pointed to heavy R&D spending, losses in the hundreds of millions of dollars, and dependence on Sol and Apollo arriving on time as risks no lab result can remove.4

The verdict

The 99.9995% figure is real, but it overstates the result: it measures memory, the tenfold gain is rounded up, and the top data point rests on three failures. The underlying result is more important than the headline. A trapped-ion machine has shown errors falling about twelvefold with each step up in code size, far faster than published superconducting experiments. If that rate holds at larger codes, fault-tolerant machines could need far fewer physical qubits than earlier estimates assumed. Whether it holds is the open question for Sol in 2027.