IBM logical-qubit results push error correction toward certified numbers

Error correction is the gate between a lab demonstration and a machine that can run a long calculation without losing its answer. If logical error rates keep falling while each logical qubit needs fewer physical qubits, hardware budgets and fault-tolerant timelines move for buyers and competitors alike. Whether that holds at large code blocks is still open.

The newest results do not settle that question, but they do shift the arithmetic for buyers.

IBM and Quantinuum report cleaner logical qubits

Quantum Zeitgeist's digest reported that IBM reached Bell-state fidelities above 93% on logical qubits through lattice surgery on ibm_phoenix, with logical errors in memory tests at 0.9% against 1.9% before. The run used 120 superconducting qubits. Much of the gain came from software: the team identified and excluded less reliable qubits and improved its decoding. Quantinuum, working with the National Quantum Computing Centre, reported a 0.26% error rate from QLDPC surgery on its H2-2 trapped-ion processor. The demonstration measured a logical qubit inside a two-qubit block using seven ancillary qubits, then reused those ancillas for repeated measurements.

Reuse matters here: every spare qubit is hardware that someone has to build and cool.

The digest also noted a 70-logical-qubit computation with verifiable results by IBM, Qedma and Algorithmiq. It reported a tenfold cut in physical qubit requirements from Google and MIT. Scaling to larger code blocks remains a challenge, the digest says. The defect-exclusion method IBM used does not scale indefinitely, the digest notes. It does make current computations more stable.

The distance to useful machines

A note published on 1 November cited a February 2026 presentation to a Japanese government quantum working group. It put current machines at 100 to several hundred qubits, error rates around 0.1% and calculations of up to 1,000 gates. Useful applications would need thousands of qubits, billions of gates and error rates of 10 to the minus 10, it says. That is seven orders of magnitude below today's error rates. By the note's reckoning, getting there would take 100,000 to 1 million physical qubits.

Set the week's results against that. A 0.9% logical error rate and a five-to-one encoding ratio beat their predecessors. They are still far from what a chemistry or materials calculation would demand.

Certified code distances enter the table

Earlier, on 30 October, Quantum Zeitgeist reported that a team at Air Force Engineering University had certified 21 new binary quantum codes. It also improved 34 others listed in Grassl's code table.

Minimum distance sets how many errors a code can absorb before the stored information is gone, and until now researchers estimated it rather than calculating it. The group tied quasi-cyclic additive codes over F4, a number system with four elements, to minimum distances it could verify. The mechanism is a theorem the team calls ancestor lifting: build a longer code from a shorter one, then check the result instead of assuming it. Divisor lattices do the bookkeeping. On a benchmark the field uses to compare codes, 21 entries now carry certified distances rather than estimates, and 34 more improved. The team also published explicit constructions for three previously unconfirmed GV codes, values that had been missing from the public record.

Six further candidates looked strong. They need exhaustive verification before anyone calls them certified, the same report says. The authors also flagged the gap between a better table and a working machine: the work does not demonstrate scalability to the sizes fault-tolerant computation would need.

Caltech and Oratomic cut the qubit count

The other half of the arithmetic is how many physical qubits each logical qubit consumes. Caltech, working with Caltech-linked startup Oratomic, said a working machine could need 10,000 to 20,000 qubits rather than the millions that earlier estimates assumed. The claim rests on an error-correction architecture for neutral atoms. Lasers called optical tweezers move and connect across a large array.

Under that scheme each logical qubit could be encoded with as few as five physical qubits, against roughly 1,000 in conventional error correction. "It's actually very surprising how well this works. It's what we call ultra-efficient error correction," said Manuel Endres, a Caltech professor of physics. John Preskill, a Caltech theoretical physicist, called the progress "optimistic". He said it brings broadly useful quantum computing closer to reality.

Caltech's figure holds only if error rates fall far enough. The lower qubit count means a machine could in theory be operational by the end of the decade, the statement said.

What happens next

Three loose ends remain. Oratomic said it will collaborate with Caltech's Advanced Quantum Computing Mission on research toward a utility-scale, fault-tolerant quantum computer. The six uncertified codes need more exhaustive verification before they join the table. Quantinuum and the National Quantum Computing Centre still have to scale QLDPC surgery to larger code blocks.

No result in this batch comes with a ship date. What changes is the arithmetic buyers and competitors use when they price a fault-tolerant system. They get fewer physical qubits per logical qubit, lower logical error rates, and code distances that can be checked rather than estimated.