Google’s Quantum ‘Magic States’ Have a 92% Rejection Rate. That’s the Point

Quantum computers need a special ingredient called a magic state. Making good ones is so costly that some estimates say it could eat up to 95% of a machine’s capacity. Researchers at Google Quantum AI now report the first successful hardware demonstration of a cheaper approach, called magic state cultivation.¹²

The work was published October 15, 2026, in the journal Science and ran on Google’s Willow superconducting processor. The states the method kept had about 40 times fewer errors than uncultivated ones. That figure covers only the attempts that passed every check, which was about 8% of them.¹²

It is a small-scale research result, and Google says it still has to be shown at larger scale. Even so, it suggests that one of the expensive chores standing between today’s machines and practical, error-corrected quantum computers can be done more cheaply on real hardware.¹

A pricey, single-use ingredient

Quantum computers need magic states to run the full range of calculations, what physicists call universal quantum computing. The catch is that each magic state is consumed when it is used.¹³

The traditional way to make high-quality ones, called distillation, takes a lot of resources. Lead author Emma Rosenfeld described the scale of the problem to Scientific American:³

“You can have up to 95 percent of your quantum computer being spent just generating these high-fidelity magic states”

Picture a bakery where nearly all the ovens make one ingredient, and only the leftover oven bakes the bread.

How cultivation works

Google researchers Craig Gidney, Noah Shutty and Cody Jones proposed cultivation in 2024.¹ In the experiment, a noisy magic state was encoded across several physical qubits, the basic units of a quantum computer. Helper qubits then scanned for mistakes without destroying the data. Any attempt that failed a check was thrown out and restarted. Only attempts that passed every check were kept.²

In other words, the method doesn’t so much repair a flawed state as refuse to keep one.

The survivors were a kind of magic state called a T-state, with a logical fidelity (how close the protected state is to perfect) of about 99.99%. Google calls that a record. “Cultivation reduces the error by a factor of 40,” the paper’s authors wrote, as quoted by Phys.org.¹²

Rosenfeld and Gidney summed it up in Google’s blog post:¹

“we’ve shown that “magic state cultivation” works in practice on our Willow processor”

Data that seemed to defy physics

Early on, a hidden measurement problem produced results that seemed to break the laws of physics. Physics was fine. Drift in electrical equipment and changes in temperature were adding noise that undermined the standard benchmark. The team added extra control and invented a new benchmark that uses cultivation to measure itself.¹

They also built a checking method called kickback tomography, which finds leftover errors with far fewer measurements than traditional techniques.² And they moved cultivated states from one error-correcting scheme (a distance-3 color code) into another (a distance-5 surface code). That proof of principle shows the cleaned-up states can survive the transfer on real hardware.¹²

What still stands in the way

Throwing out roughly 92% of attempts has a cost. Secondary summaries of Scientific American’s reporting say Kenneth Brown of the Duke Quantum Center called that rejection rate a meaningful concern, while accepting that cultivation works in practice. The same summaries put the kept states’ error rate at roughly 1 in 10,000, still far above what demanding jobs like simulating molecules require.⁵

The Quantum Insider framed the result as an important but incremental step that doesn’t necessarily solve the resource problem at practical scale. Larger tests and better ways to store and use the states are expected to be needed.⁴ Google itself says cultivation must be demonstrated at larger scale to be realistic for practical quantum computing, and that plenty of technical challenges remain.¹