Quantinuum Helios Sets Error Correction Record at 99.9995%

What happened

Quantinuum announced on October 19, 2026, that its Helios trapped-ion computer reached a logical fidelity of 99.9995%. That works out to about five errors per million operations on error-corrected qubits.1119 The company says this error rate is 10 times lower than the record it set one month earlier. It also puts the figure 100 times below the best published result from any other hardware type, a superconducting result from Google Quantum AI.14 Quantinuum's CEO, Rajeeb Hazra, said the result shows the company has the error-correction design and trapped-ion technology needed for universal fault tolerance.14

The machine has 98 physical qubits. Each one is an ion controlled with lasers and electromagnetic fields.12 To build the logical qubits, the team used a well-studied scheme called a color code.12 Outside experts were positive. Robin Blume-Kohout, co-director of the Quantum Performance Lab at Sandia National Laboratories, called it a milestone researchers have chased for years. He pointed to a logical qubit that performs nearly 100 times better than the physical qubits it is built from.4 Steven Flammia of Virginia Tech, who was also not involved, described it to New Scientist as an impressive technical advance.12

The headline number is real, and the scaling behind it matters more than the record itself. But the "100x better than Google" framing used across much of the coverage needs some unpacking.

The scaling result matters most

The most important data point is not the 99.9995% figure. It is how errors changed as the code got bigger. Quantinuum tested three code distances, 3, 5 and 7. Each step up cut the remaining logical memory errors by an average factor of 12, and logical gate errors by an average factor of 9.13 New Scientist put it in hardware terms: errors fell twelvefold when a logical qubit grew from seven to 19 physical qubits, and fell by the same factor again from 19 to 37.12

This matters because error correction only pays off if bigger codes actually work better. A system that improves with each added layer of protection can, in principle, keep suppressing errors as it grows. A system that plateaus hits a ceiling. One analysis notes that a steady factor-of-12 suppression is evidence of exponential error reduction, not a one-off tuned demonstration. That is a big part of why the result reads as more than marketing.13 Blume-Kohout said the findings go a long way toward retiring one of the field's central technical risks. He added that a scaled-up version could reach useful fault-tolerant reliability with fewer than a million physical qubits.1

According to the team's lead researcher, David Hayes, part of the gain came from better control of the physical qubits. When a qubit's energy drifts in a harmful way, a corrective laser pulse now pushes it back, and the color code then handles the remaining errors.12

Same hardware, different code

The October result ran on the same Helios machine as the September record. The only change was the error-correction code.19 In September, Quantinuum used a code called Helix, which needed about 10 physical qubits per logical qubit. The color code needed up to 37 but produced lower error rates.11 The company presents this as proof of a flexible design that can be retuned for accuracy, efficiency or speed without new hardware.11 The architecture adjusts its error-checking measurements on the fly and links codes suited to different tasks. It also uses real-time decoding, including AI decoders built for the system.144

The overhead trade-off deserves attention. A code that uses nearly four times the qubits for a 10x accuracy gain may be worth it for some workloads and not others. Having both options on one machine is a practical advantage. It is also a reminder that "logical qubit" covers a wide range of resource costs.

Unpacking the comparison with Google

The reporting agrees on the basic numbers. Where it differs is in how much context it gives about what was measured. The 99.9995% figure is fidelity per error-checking cycle on a protected memory qubit, at distance 7. That is a failure probability of about 5 in a million per cycle.13 TipRanks described it as five errors per million error-checking cycles, which is more precise than the "operations" wording in most headlines.6

The Google benchmark is a distance-7 surface-code result of roughly 1.43 errors per thousand cycles.13 Quantinuum's release links it to a Nature paper with a 2024 identifier, which points to Google's earlier Willow-era work rather than anything from this month.20 Google's own roadmap still lists its current logical error rate at about 10^-3 per cycle. Its target is 10^-6 for a long-lived logical qubit, a milestone it places around 2027.16 Quantinuum is now close to the error level Google has set as a future goal. But the comparison pits a 2026 trapped-ion memory result against a superconducting result published roughly two years earlier.

Google has not been idle. The week before Quantinuum's announcement, Google Quantum AI showed a technique called magic state cultivation that cut errors by a factor of 40.9 Magic states are a key ingredient for universal computation, which the Quantinuum release does not claim. One analysis notes that the October 19 announcement covers logical memory and scaling, not full universal computation with magic-state injection at this fidelity.13

Computation is harder than memory. When the team ran a simple error-corrected operation on two logical qubits, the error rate rose about tenfold, to roughly seven errors per 100,000 uses.12 That is still strong. But it is a different number from the one in the headlines, and it is the one that bears more directly on running real programs.

Roadmaps and the logical-qubit race

Quantinuum's roadmap runs from Helios to a system called Sol in 2027, then to Apollo, a universal fault-tolerant machine targeted for 2029.6 The company says its accuracy is now within Apollo's target range.19 It estimates that 100 protected logical qubits would need about 30,000 physical qubits. Its comparison model for superconducting hardware needs about 900,000.3

The numbers on logical-qubit counts are more modest. One tracker that counts only fully error-corrected results, with no failed runs thrown away, lists Quantinuum's record as eight logical qubits, set in September. That result came from one shallow benchmark circuit.16 On those strict terms, rivals such as Google, IBM and QuEra are credited with two logical qubits each.16 Several companies have announced larger counts, but they rely on codes that can only detect errors or on discarding runs where errors were flagged. That approach is useful but is not the same as correction.16

The 2029 timeline is getting crowded. Google's roadmap aims for a useful error-corrected computer with about a million physical qubits and 1,000 logical qubits. IonQ has published a goal of 8,000 logical qubits for the same year.16 The same tracker notes that quantum roadmap dates slip more often than they hold.16 Quantinuum's own forward-looking statements also warn that its results were achieved in experimental settings and may not carry over to larger systems or commercial use.20

Why it matters

The result supports a view that has been gaining ground: the quality of physical operations may matter more than raw qubit counts. Trapped-ion systems have generally led on gate fidelity, while Google and IBM have pushed for scale.3 If a 100-fold gap in logical error rates holds up under independent checking, it strengthens the case for quality first.3 One analysis argues that the bigger implication is cost. Reaching fault tolerance with tens of thousands of physical qubits instead of hundreds of thousands would sharply cut the capital and engineering needed.3

The result has also drawn attention from cryptocurrency observers, who track how fast quantum machines are approaching the point where they could threaten the encryption behind bitcoin.9 That concern is still far off. Breaking RSA-2048 is estimated to need about 4,000 logical qubits, hundreds of times beyond today's counted record.16

The business side of the story is less advanced than the science. Quantinuum lost $65.4 million in its latest quarter on revenue of about $8 million.8 It is expanding access to Helios through Oracle's cloud platform.6

The bottom line

This is a substantive milestone. Steady factor-of-12 error suppression on a commercial machine, confirmed by respected outside experts, is the kind of evidence the field has been waiting for. But the "five errors per million operations" figure describes memory cycles, not full computation, and the Google comparison is against a result that is roughly two years old. The next real tests are reproducing these error rates during logical computation at useful depth, adding magic states, and hitting the Sol and Apollo dates.313