Quantinuum Error Correction Record Hits 99.9995% Fidelity on Helios

A second record in about a month

On October 19, Quantinuum said its Helios trapped-ion computer reached a logical fidelity of 99.9995%. The company calls this a world record for quantum error correction, and it works out to roughly five errors per million operations.1 The result came from a well-known scheme called a "color code," run inside the company's own error-correction architecture. Quantinuum says the error rate is ten times lower than the record it set in September and about a hundred times lower than the best published result on any other hardware, which belongs to Google Quantum AI's superconducting system.1

The September record used a different code, Helix, on the same Helios machine.18 Quantinuum is using that back-to-back pairing as evidence that its architecture can switch codes and keep improving without a hardware upgrade. CEO Rajeeb Hazra said the company has now proven it has the architecture and trapped-ion technology needed for universal fault tolerance.1

The 99.9995% figure gets most of the attention, but the scaling behind it matters more.

The number that matters is 12

Error correction groups many unreliable physical qubits into one sturdier "logical" qubit. It only pays off if the physical hardware is good enough. Above a certain error threshold, adding qubits makes results worse. Below it, each larger code should cut logical errors exponentially.8 The size of that gain is measured by a suppression factor, usually written Λ (lambda).

Quantinuum tested color codes at distances 3, 5 and 7. Each step up cut the remaining logical memory errors by about a factor of 12 on average, and cut logical gate errors by about a factor of 9.7 A detailed breakdown of the paper puts the logical error per round at 5.5 × 10⁻⁴ at distance 3, 3.8 × 10⁻⁵ at distance 5 and 5 × 10⁻⁶ at distance 7, with a fitted Λ of 12.4

The same analysis lists earlier results for comparison: 2.14 for Google's 2024 surface-code experiment on Willow, 1.56 for Google's 2025 color-code run, and 1.40 for USTC's Zuchongzhi 3.2.4 Even allowing for the experiment's error bars, which range from roughly 10 to 15, the gap is large. Most of the field's best results so far showed suppression just above 1, which means barely below threshold. Helios appears to be operating well below it.

The reason is partly physical. Helios's average two-qubit gate error is about 8.4 × 10⁻⁴ with a new "leakage repump" technique switched on.4 That fix optically returns barium ions that have drifted out of the computational states, with no added circuit time.8 The company says the repump improved its logical error rate by more than an order of magnitude and made the distance-scaling demonstration possible.7 Helios's all-to-all connectivity also matters, because it lets ions be reordered through software relabeling during the color code's error checks.8

Read the fine print on "five errors per million"

The sources agree on the headline figures but describe them differently, and the differences are worth noticing.

First, the 99.9995% is a fidelity per round of error correction for a stored logical qubit, not for a running computation. The team prepared a logical state, ran 20 rounds of error checks and measured it, with no computation in between.4 Press releases and general-interest headlines described this as "operations" or even "accuracy," which makes it sound broader than it is.6 The paper's figure for actual computation is a logical two-qubit gate error of 7 × 10⁻⁵ at distance 5. That is 14 times better than the same gates on bare physical qubits, which is impressive but not five-per-million.4

Second, the qubit cost is reported inconsistently. The company's release says Helix used 10 physical qubits per logical qubit, against "up to 37" for the color code.1 Detailed technical coverage gives 13, 37 and 73 physical qubits per logical qubit at distances 3, 5 and 7, counting both data and measurement qubits.48 The record-setting distance-7 run therefore appears to have needed 73 qubits, not 37. That gap is significant for a company whose pitch is qubit efficiency.

Third, the size of the lead over Google changes with the comparison. The release says a hundredfold.1 Independent tallies put Google's distance-7 surface-code result at about 1.43 × 10⁻³ per cycle.5 Quantinuum's own blog cites a different Google figure, 0.0772% per cycle at distance 7, and claims a 154x improvement.7 Either way the lead is large, but the exact multiple depends on which baseline you pick.

Fourth, speed. Trapped ions are slow. One analysis puts each correction round at about a fifth of a second and argues that speed, not accuracy, is now the main constraint.4 That analysis also notes that, even per second, a stored Helios logical qubit at this error rate would fail only about once every 11 hours.4 Accuracy per unit time is not the only thing that matters, though. Useful algorithms need billions of operations, and wall-clock time is a real problem for ion-trap systems.

The outside view, and how it got stretched

The most quoted outside comment came from Robin Blume-Kohout, co-director of the Quantum Performance Lab at Sandia National Laboratories, who was not involved in the study. He called it a milestone researchers had worked toward for years, said the logical qubit performs nearly 100 times better than its physical parts, and said a scaled-up version could reach useful fault tolerance with fewer than a million physical qubits.2 He added that the work goes a long way toward retiring one of the field's central technical risks.13

That is a strong endorsement, but it is phrased as conditional. Some coverage went further: one outlet said Sandia had "confirmed" sub-million-qubit fault tolerance and called the Apollo machine "feasible."8 What Blume-Kohout actually said is that a scaled-up system could get there.2 More careful reporting noted that Quantinuum's claim of needing about 30 times fewer physical qubits than leading superconducting approaches is a projection from the measured suppression factor, not a demonstrated system size.5 That projection extrapolates the Helios curve to a logical error rate of 10⁻¹² at code distance 21, compared with distance 109 for Google's color-code curve.4

Roadmap: Helios, Sol, Apollo

The commercial stakes are about timelines. Quantinuum says its accuracy is now within the target range for Apollo, the universal fault-tolerant computer it plans to make available by 2029.18 An intermediate system, Sol, is expected around 2027.8 Helios itself is a 98-qubit machine.6 It is already sold commercially and, as of August, offered as an Oracle Cloud service.8

Some pieces are still missing. Earlier Quantinuum work on code switching produced high-fidelity logical T-state preparation, but has not yet been combined with the below-threshold color-code scaling shown here.8 T-states, also called "magic states," are needed for universal computation, so combining the two is the next milestone to watch.8

The error budget also shows where future gains have to come from. Two-qubit gate error accounts for 43% of logical error, memory and transport for 36%, and measurement crosstalk for 13%.8 Better hardware gates would raise Λ further. Scaling from 98 qubits to the thousands that Apollo-class logical counts need is a separate engineering problem, and this result does not solve it.

Where this fits in a crowded field

The main metric has been shifting. Over the past two years, companies have competed on how many logical qubits they can show: QuEra reported 96 error-corrected logical qubits in January 2026, and Quantinuum showed 48 error-detected logical qubits on Helios in November 2025.9 Those counts measure breadth. The October result is about depth: how good one logical qubit can get, and how fast it improves as the code grows. Fault tolerance needs both, and no company has yet shown both at scale on one machine.

Investors were less impressed. Quantinuum shares fell about 7.6% around the announcement.14 Financial coverage noted heavy R&D spending, losses in the hundreds of millions of dollars, and dependence on Sol and Apollo arriving on time.14 The company reports third-quarter results on November 5.11

The verdict

The press-release language overstates some details. "Five errors per million operations" describes memory, not computation. The record run likely used 73 physical qubits, not 37. The hundredfold lead depends on which Google baseline you use. Even so, this is a substantive milestone, not just bragging. A suppression factor around 12 on commercial hardware, backed by a credible outside expert, is a qualitative change. It suggests error correction on trapped ions has moved from barely working to clearly working. The open questions are now speed, scale and magic states, not whether error correction works. That is a much better position for Quantinuum to be in than a year ago, but it is still not a fault-tolerant computer.