Google cut quantum errors 40x on Willow with a trick called magic state cultivation

Google Quantum AI says it has shown that “magic state cultivation” works on real hardware. According to the official announcement, the technique cut errors about 40x on the company’s Willow processor. The results appear in the journal Science.

What Google demonstrated

Magic states are special quantum states that a quantum computer needs in order to run the full set of operations behind useful programs. Google describes them as a long-standing bottleneck, one researchers have anticipated since 2004. Producing them with error correction is very resource-heavy, and a magic state cannot be reused once it has been consumed. Google cites estimates that up to 95% of a quantum computer’s capacity might have to go to making them, leaving little room for actual computation.

Cultivation, proposed in 2024 by Google researchers Craig Gidney, Noah Shutty and Cody Jones, works like a filter. A noisy state is injected into a small region of a chip, and a series of coordinated measurements checks it for errors and discards flawed attempts. What remains is a very pure, low-error state that can be adapted into a larger encoding.

On Willow, Google reports a 40-fold error reduction compared with states that were not cultivated. The team also produced T-states, one common type of magic state, with a record logical fidelity of 99.99%.

The team also tested whether the theory fits Google’s architecture. It moved states between two kinds of error-correcting code: color codes, which Google calls well suited to cultivation, and the more established surface codes. Specifically, it grafted states from a distance-3 color code onto a distance-5 surface code. According to Google, this showed that purified states can survive the transfer inside the hardware.

The measurement problem along the way

Google’s account of the work includes a detour. Research scientist Emma Rosenfeld first saw data that appeared to contradict the laws of physics. After months of simulations and consistency checks, she traced it to a wrong assumption in the team’s standard measurement method. Noise from electrical equipment drift and temperature variation was undermining the benchmark.

The team responded by adding control to reduce the effect of that noise. It also devised a new benchmarking approach that uses cultivation to measure itself. Google says this set a new state of the art in quantum metrology.

Why it matters

The practical question is how much hardware cultivation could save, and published estimates give a sense of scale. In the original proposal, Craig Gidney, Noah Shutty and Cody Jones reported on arXiv that cultivation needs an order of magnitude fewer qubit-rounds than prior approaches to reach low logical error rates, and roughly the same number of physical gates as a lattice-surgery CNOT of equivalent reliability. Those figures come from simulations at an assumed noise level of 0.1%, with logical error rates as low as 2 in a billion. The Willow result tests the underlying mechanism on a chip, not that full cost claim.

The savings flow straight into machine size. In a 2025 paper, Gidney estimated that a 2048-bit RSA integer could be factored in under a week with fewer than a million noisy qubits, against a 2019 estimate of 20 million qubits and eight hours. He credits several changes, including allocating less space to magic state distillation by using cultivation. The trade-off is time: the paper notes that runtime grows partly because fewer magic state factories are used.

That makes the economics a three-way balance among qubit count, runtime and error rate, and cultivation moves the balance toward smaller machines. These are architecture estimates built on assumptions about hardware quality, not forecasts of what a built machine will cost. Newer proposals using different codes claim further cuts, but they change the assumptions and are not directly comparable. Whether Willow-scale results hold up in larger codes is still the open question.

What’s next

Google says cultivation still has to be shown at a larger scale before it is realistic for practical quantum computing, and that plenty of technical challenges remain. The stated goal of Google Quantum AI is a useful, error-corrected quantum computer that can solve problems that are otherwise out of reach. The paper is available in Science, and Google did not give a timeline for larger demonstrations in the announcement.