What is a cat qubit in quantum computing?

It is named after a famous thought experiment by the physicist Schrödinger. In it, a cat in a box, hooked up to a device that follows the rules of quantum physics, ends up in a superposition of being alive and being dead. A superposition of two clearly different states is called a “cat state,” after this story.

A cat qubit uses two states of the light in a box (a resonator) that swing in exactly opposite ways. Each is close to ordinary light, with a set strength and a set timing of its swing. A cat state, the superposition of the two, is a superposition of two clearly different states, like a living cat and a dead cat. No real cat is used anywhere.

Inside a circuit made of superconducting metal (metal whose electrical resistance drops to zero when it is cooled enough), there is a resonator that acts as a tiny box. A resonator is a part that can hold, for a while, a kind of radio wave that swings at a set rate (microwaves). AWS’s 2025 chip used resonators shaped like thin lines drawn on a flat plate as the boxes for its cat qubits.

A cat qubit takes two states of the light held in this box, swinging in exactly opposite ways, as 0 and 1: when one wave is at a crest, the other is at a trough. Because it is a qubit, it can also be in a superposition of 0 and 1. Overlapping 0 and 1 half and half with a plus gives one cat state, and with a minus, the other.

The more particles of light (photons) in the box, the more clearly the two states differ, and the harder they are to mix up. In the 2025 experiment, the average number of photons was set between 1 and 4.

For 0 to turn into 1, the way the light in the box swings has to reverse completely. When a small disturbance makes a single photon escape or enter, the swing doesn’t reverse completely. On top of that, a special circuit that puts photons into the box and takes them out two at a time keeps pulling the light back to those two states (and their superpositions). This way of keeping them was proposed in a 2014 paper (Mirrahimi et al.). The more photons, the farther apart the two states are, so each extra photon makes bit flips several times rarer.

In exchange, phase flips increase. Every time a single photon escapes from the box, the plus and minus in how 0 and 1 overlap swap. With more photons, more escape, so how often phase flips happen grows roughly in proportion to the number of photons.

In AWS’s 2025 experiment, with an average of 2 photons, the time until a bit flip was more than 1 millisecond (a thousandth of a second), and until a phase flip, 27 to 33 microseconds (about 30 millionths of a second). That makes bit flips more than 30 times rarer. Errors this lopsided toward one kind are the cat qubit’s strength.

Cat qubits are lined up in a row, and the whole row holds one piece of information. One piece of information held by many parts like this is called a logical qubit. Between each pair of neighboring cat qubits sits a checker qubit (an ordinary superconducting qubit). A checker never reads the contents; it only checks whether the two neighbors are the same. From where they disagree, a calculation estimates which cat qubit likely had a phase flip. This arrangement is called a repetition code.

A repetition code can fix only one kind of error. To fix both kinds on ordinary qubits, schemes such as the surface code, which lays qubits out on a flat grid, are used, and they need many parts. Because bit flips are so rare in cat qubits, one row is enough. The idea of using cat qubits in a one-row repetition code was laid out in a 2019 paper (Guillaud and Mirrahimi).

AWS says a row of five (a code of distance 5) needs only 9 qubits, 5 cat qubits and 4 checkers. The distance of a code is a measure of how many overlapping errors it can still sort out; in a one-row code, it is the number of cat qubits in the row. A surface code of the same distance needs 49, so this is less than a fifth. On top of these, there are 5 circuits that keep the cat qubits in their two states.

In February 2025, researchers at AWS (Amazon Web Services) published a paper in Nature. AWS announced the chip under the name Ocelot. On the chip, they ran an experiment storing one piece of information in a row of five cat qubits (distance 5) and in two rows of three within it that overlap as little as possible (distance 3).

Phase flips went down when the row grew from three to five, at every photon number they tested. Adding parts only cuts the overall errors when each part’s own errors are low enough, so for phase flips, the parts met that condition. Bit flips, on the other hand, rose in the longer row. A one-row code can’t fix bit flips, so a bit flip in any one cat qubit of the row spoils the whole row’s information, and a longer row has more places where that can happen. Even so, bit flips fell as photons were added, and even in the row of five, at an average of 4 photons, they dropped below 1% per round of checking.

Counting both kinds, the lowest leftover error per round of checking was 1.75% on average for the rows of three and 1.65% for the row of five (for the row of five, at an average of about 1.5 photons). The paper says the row of five, despite having more places where errors can happen, kept its errors about as low as the rows of three. One round of checking took about 2.8 microseconds.

But practical use is still a way off. Most of today’s errors come from the cat qubits’ own bit flips and phase flips. And because a repetition code can’t fix bit flips, making the row longer and longer stops cutting errors at some point. The paper says bit flips caused by the checkers will eventually need fixing too, which could be done by building cat qubits into surface codes tailored to lopsided errors. AWS says that scaling this approach up could cut the cost of fixing errors by up to 90% compared with current approaches, but that is the company’s projection, and Ocelot is still a prototype.

In October 2026, researchers at Alice & Bob, a Paris company that builds cat qubits, estimated what it would take to break RSA-2048, one of the encryption methods widely used today, with a quantum computer. It is a paper not yet checked by other researchers, and the company’s own estimate: about 23,000 cat qubits (22,593) and about 3.14 days.

The number rests on assumptions: one round of checking takes 1 microsecond, deciding the next step from the checks takes 10 microseconds, phase flips happen at a rate of 1 in 2,000 for each operation in the circuit, and bit flips don’t happen at all (equal to about 100 years until a bit flip). If bit flips happen about once an hour, a small code that fixes bit flips has to be added, and the paper says the count rises to about 45,000.

The times until a bit flip confirmed in papers so far are more than 1 millisecond for the cat qubits on AWS’s 2025 chip, and more than 10 seconds for a single cat qubit in 2024. The estimate assumes far longer times than these. One round of checking also took about 2.8 microseconds in the 2025 experiment, longer than the assumed 1 microsecond.

For ordinary qubits with the surface code, Craig Gidney of Google Quantum AI estimated in 2025 that it could be done with fewer than a million qubits in less than a week (assuming a 0.1% error per operation, 1 microsecond per round of checking and so on). The kinds of parts and the assumptions differ, so the numbers can’t be compared directly. Both are estimates for machines that haven’t been built yet.