On September 24, 2026, Infleqtion announced that it had run a circuit handling 30 logical qubits using 80 physical qubits on its neutral-atom quantum computer, Sqale. The result extends, to 30 logical qubits on the company's own hardware, a scheme that spreads quantum information across multiple atoms so that errors can be detected during computation.
At this point, however, the results can be checked mainly through the company's announcement and a technical write-up. A detailed paper is expected to be published within the next few weeks.
Building 30 logical qubits from 80 physical qubits and running a circuit on them is a different stage from sustaining a long quantum computation without errors. It is worth separating what this experiment demonstrated from what remains open.
30 logical qubits built from 80 atoms
In the experiment, the 80 atoms were divided into 10 groups of 8, and each group encoded 3 logical qubits.
The experiment itself was completed in August. To convert the quantum circuit into a form that could run on Sqale hardware, the team used the company's software, Superstaq. The results are not from simulation alone.
The setup and experimental conditions are described in a technical explanation by CTO Pranav Gokhale.
A logical qubit is a mechanism that holds a single piece of quantum information distributed across multiple physical qubits.
Using extra physical qubits makes it possible to carry information for checking whether an error has occurred. However, which types of errors can be detected or corrected depends on the quantum error-correcting code used.
For that reason, the ratio of physical to logical qubits alone cannot tell you how strongly the logical qubits are protected.
Sqale uses a distance-3 [[8,3,3]] code to prepare the initial state, and a distance-2 [[8,3,2]] code for the subsequent operations.
The notation gives, in order, the number of physical qubits used, the number of logical qubits encoded, and the code distance.
A distance-2 code can detect an error on any single qubit. That does not mean it can correct every single-qubit error it detects.
It therefore cannot be concluded that the computation was protected by distance-3 error correction throughout, just because a distance-3 code was used for state preparation.
One of Sqale's features is that, besides moving atoms to change which ones connect to each other, it can also operate on selected atoms without moving them.
Rich Rines and colleagues at the company previously reported experiments including one that handled 12 logical qubits using 114 physical qubits, in the preprint "Demonstration of a Logical Architecture Uniting Motion and In-Place Entanglement."
The first version was posted in September 2025 and a revised second version in April 2026. It is separate from the 30-logical-qubit experiment announced now.
About 25% of outputs landed in the set that an ideal circuit can produce
The circuit run on the 30 logical qubits is of a type called IQP.
This time, each group was prepared in the |+++> superposition state, a phase-shifting operation was applied, and the qubits were then measured in the X basis.
The circuit contained four logical CCZ gates and, in total, about 1,000 physical operations. CCZ is a kind of non-Clifford gate acting on three qubits, and an important operation for building universal quantum computation.
The metric Gokhale presented is what percentage of the measured outputs fell within the set of outputs that an ideal circuit could produce.
There are 2³⁰ possible outputs from 30 bits. The outputs that can appear in the ideal circuit are limited to 262,144.
According to Infleqtion, roughly 25% of the outputs obtained in the experiment were contained in this set.
If all 30-bit strings were output with equal probability at random, the chance of landing in this set by accident would be
262,144 ÷ 2³⁰ = 1/4,096 ≈ 0.0244%.
Comparing this roughly 0.0244% with the experiment's roughly 25% gives a difference of about 1,000 times. This is what Infleqtion's "about 1,000x" refers to.
However, this does not mean computation became 1,000 times faster, nor that the correct-answer rate on a practical problem is 1,000 times higher.
The baseline is "perfectly random output that produces every 30-bit string with equal probability." It is not a comparison with the best classical algorithm, nor with running the same circuit without quantum error correction.
Furthermore, the fraction that landed in the correct set does not reveal how closely the individual outputs within that set matched the ideal probability distribution.
Even at the same 25%, the distribution inside the set could differ substantially from that of an ideal quantum circuit.
The roughly 25% therefore cannot be treated as the fidelity of the whole circuit, nor can this number alone be used to judge that quantum advantage has been demonstrated.
The published technical explanation does not state the total number of shots or the acceptance rate when post-selection is applied, and it gives no statistical error range for the roughly 25% figure.
It is also unclear whether the roughly 25% was calculated before or after the process of restoring values missing because of atom loss.
How many runs were performed, and how consistently comparable results can be reproduced, will need to be confirmed in the forthcoming detailed paper.
Measurements from lost atoms are restored after the computation
In the Sqale experiment, even if one atom is lost within a group of eight, the missing X-basis measurement value can in some cases be restored afterward by software.
Gokhale gives as the reason that "when an atom is lost, it is possible to identify which qubit was lost."
In an "erasure error," where the location of the lost information is known, more information is available than with a typical qubit error whose location is unknown.
The restoration uses even/odd relationships that combinations of measurement values must satisfy, that is, parity constraints.
If the position of the lost qubit is known, the missing value can sometimes be inferred from the remaining measurement results and this constraint.
While this processing increases the number of runs usable for analysis, Infleqtion acknowledges that the error rate also becomes higher.
However, the currently published explanation does not say how many usable results there were before and after the processing.
The key point is that this is data processing after measurement is finished.
It differs from full-fledged quantum error correction, in which errors are measured repeatedly during the computation and control is applied in real time according to the results as the computation continues.
Infleqtion itself lists real-time control and mid-circuit measurement as future priorities.
The handling of atom loss in this experiment therefore cannot be taken to mean that continuous quantum error correction has been achieved.
AI helped cut the gate count for one operation from 8 to 4
The team also worked to run the quantum circuit with as few operations as possible.
According to Infleqtion, a logical operation called "double-CZ," found with help from the AI model "GPT 5.6 Sol," halved the number of physical two-qubit gates from 8 in the comparison implementation to 4.
The circuit as a whole used 136 physical two-qubit gates.
However, the reduction from 8 to 4 applies to a specific logical operation. It does not mean that every quantum operation in the circuit was halved.
Infleqtion explains that reducing the gate count also improved the proportion of valid results obtained and the computational accuracy.
The published technical explanation, on the other hand, does not include data directly comparing accuracy or success rate before and after using double-CZ.
The fact that fewer gates were needed and the degree to which that improved actual error rates need to be evaluated separately.
30 logical qubits is not a world record
There is earlier research on neutral-atom quantum computing with logical qubits that handled more logical qubits than this one.
In a peer-reviewed Nature paper published online in December 2023, Dolev Bluvstein and colleagues reported circuits using up to 48 logical qubits.
That work also used the [[8,3,2]] code, and incorporated 228 logical two-qubit gates and 48 logical CCZ gates into a sampling circuit with up to 48 logical qubits.
It also compared circuits using quantum error-correcting codes with circuits using only physical qubits, using a metric called cross-entropy, to examine how much error detection improves results.
Because of this earlier work, Infleqtion's 30 logical qubits cannot be called a "world record for the number of logical qubits."
The "first" claims in Infleqtion's announcement come with qualifiers such as "neutral-atom quantum computer company" and "commercial system."
It is more accurate to see this result not as a simple record for the most logical qubits, but as progress in running a circuit at the scale of 30 logical qubits on real hardware, using Infleqtion's own device and control methods.
Because circuit types, error-correction schemes, and gate counts differ, the number of logical qubits alone also cannot be used to rank the performance of different machines.
What matters next is less "how many" than "how long and how accurately"
New proposals have also emerged on how logical qubits should be evaluated.
In the non-peer-reviewed preprint "Scalable logical qubits," submitted September 17, Matthias Troyer, Chetan Nayak, and John Martinis proposed a framework that evaluates logical qubits not only by their reliability and scale but also by the kinds of operations they can run and their processing performance.
They place particular emphasis on being able to continue computing while repeatedly performing error correction, and on real-time decoding, which analyzes detected errors with low latency and feeds the result into the next control step.
The authors point out that with post-selection, which keeps only runs in which no error was detected, the fraction of usable trials could fall rapidly as computations get longer.
This is a proposal about how to evaluate logical qubits, not an established industry standard. Nor is it independent verification of the Sqale experiment.
Still, it shows the need to look beyond the number of logical qubits to how long a circuit can be run and what percentage of the results can actually be used.
The units of operation counts also call for care.
The "about 1,000" here is the number of physical operations performed on physical qubits.
The roughly one million that Infleqtion aims for in the future is a target for logical operations protected by quantum error correction, so the two do not count the same thing.
John Preskill's "Beyond NISQ: The Megaquop Machine," published in ACM Transactions on Quantum Computing, also discusses the scale of performing more than a million logical operations on a machine protected by quantum error correction.
One therefore cannot simply divide the roughly 1,000 physical operations by one million logical operations and say what percentage of the way to practical use has been covered.
As of September 28, the public materials available include no peer-reviewed paper corresponding to this 30-logical-qubit experiment and no replication by an independent research team.
Infleqtion has set a goal of moving to 100 logical qubits in 2028, but the number of logical qubits is not the only thing that matters in assessing that progress.
How much more can errors be suppressed than with physical qubits when error detection and correction are repeated during computation? Can a sufficient fraction of usable results be retained even when long circuits are run? And how far can computation continue while real-time error correction is performed?
Once such data are in hand, it will be possible to judge not just the number of logical qubits but how close the machine has come to a quantum computer that can carry actual computations through to the end at the required accuracy.