Google shows magic state cultivation works on its Willow quantum chip, cutting errors about 40x

Google Quantum AI says it has shown, on real hardware, a technique that has been a theoretical promise for two years: "magic state cultivation." In a paper published in Science on 15 October 2026, the team reports that running the protocol on its Willow processor cut errors in a key kind of quantum state by about 40 times compared with preparing the same state without cultivation. It is a narrow result, but it targets what many researchers regard as one of the most expensive line items in building a useful quantum computer.

What happened

On 15 October, Google published a blog post by research scientist Emma Rosenfeld and software engineer Craig Gidney, alongside the Science paper "Magic state cultivation on a superconducting quantum processor" (DOI 10.1126/science.aee7961). The headline figures, as Google reports them:

• Errors in the prepared states fell about 40-fold relative to states made without cultivation.
• The resulting "T-states" reached a logical fidelity of 99.99%, which Google calls a record. The Quantum Computing Report gives the figure as 99.99(1)%, an error rate of about one in ten thousand.
• The experiments ran on Willow, which The Quantum Insider describes as a 105-qubit superconducting chip on a square lattice.
• The team also moved a purified state from a distance-3 color code into a distance-5 surface code, a step called "grafting." The Quantum Insider says the distance-5 experiment used 54 qubits.

Google is careful about its own claim: "we still need to demonstrate magic state cultivation at a larger scale" before it is realistic for practical machines, the company says, and "plenty of technical challenges remain."

Background: why magic states matter

Quantum computers protect their data with error-correcting codes, which spread one logical qubit over many physical qubits. Some operations are easy to perform on protected data. Others, notably the "T gate" needed to run programs a classical computer could not simulate efficiently, are not. The standard workaround is to prepare a special resource called a magic state, then consume it to perform the hard operation. A magic state is used up in the process, so a large program needs a steady supply.

Making them clean enough is the costly part. Google's post says researchers have foreseen this bottleneck since 2004 and that some estimates put up to 95% of a quantum computer's capacity on producing magic states, leaving little for the computation itself. The conventional route is "distillation": make many noisy states and combine them into fewer, cleaner ones. That works, but it is hungry for qubits and time.

In 2024, Google researchers Craig Gidney, Noah Shutty and Cody Jones published the cultivation proposal, which Google says refined an idea from earlier literature. Instead of distilling many states, you inject one noisy state into a small patch of the chip and run a series of coordinated checks. If a check detects an error, you throw the attempt away and start over; the survivors are very clean. The cleaned state can then be switched into the larger code that the rest of the machine uses. On paper it promised a large cut in the hardware needed. Until now it had been shown in theory and simulation, not on a physical device.

How it worked on Willow

Three details from the paper coverage show where the engineering effort went.

Filtering means discarding. The Quantum Computing Report says that after fault-tolerant measurement filtering, 8% of attempts were retained. That matters for cost: a state that is very clean but survives rarely still needs many tries. The Quantum Insider notes that the retained fraction matters alongside fidelity when estimating computing cost, and that its article gives no rate itself. We cite the 8% figure to the Quantum Computing Report only.

Calibration with machine learning. According to The Quantum Insider, the team used reinforcement learning as a final calibration step, aimed at reducing how often the circuit flagged errors. For one run at roughly 1-in-10,000 error, training took about four hours before data collection began. The outlet reports this improved both the logical error probability and the retained fraction by about an order of magnitude, and that repeated experiments, including fresh training runs, agreed within statistical uncertainty.

A measurement trap. Google's post describes months in which Rosenfeld saw data that "seemed to defy the laws of physics." The cause was an incorrect assumption in the team's standard measurement technique: low-frequency noise, which the post links to equipment drift and temperature variation, made the benchmark unreliable. The Quantum Insider adds that under such noise, the usual way of combining measurements could keep different subsets of attempts and produce estimates that were not physically possible, and could also overstate fidelity without an obviously impossible result. The fix was a new technique, which the paper calls "kickback tomography" according to that outlet, that uses cultivation to measure itself and reports a lower bound on fidelity. Reporting a bound rather than a point estimate is a conservative choice: the true fidelity should be at least as good.

What it means

For hardware teams, the value is the resource arithmetic. If a magic state can be grown in a small patch rather than in a large distillation factory, the physical machine needed to run a given algorithm shrinks. Google says as much: it expects cultivation to substantially reduce the size of the quantum computer needed for practical tasks. That is a projection, not something this experiment measured.

For everyone else, it is a reminder of how the field makes progress. The headline milestones, such as qubit counts, get attention, but a lot of the path to a useful machine runs through unglamorous pieces like this one. A 40x improvement on one subroutine does not make Willow 40 times better, and Google does not claim it does.

What we don't know yet

• Scale. The grafting demo was into a distance-5 surface code. Practical algorithms need much larger codes and many states produced in parallel.
• Comparison with distillation. The 40x figure compares cultivated and uncultivated states, not cultivation against distillation on the same hardware. A reader should not take it as a head-to-head result.
• Storage. The Quantum Insider reports that when the team ran extra error-correction cycles after grafting, the logical error rate was about seven times higher than simulations predicted. The authors point to leakage, where qubits stray outside the two computational states, as the likely main cause: the simulations left it out and the circuits had no leakage-removal operations. Planned fixes include better calibration, leakage removal, simpler circuits and larger codes.
• Independent replication. We found no outside expert commentary in the coverage we read; both secondary outlets repeat Google's account. The paper has been through Science's review, but other groups have not yet reproduced it.
• Our reading. We read the Google post and secondary coverage, not the full paper text, so details such as exact error bars come from those summaries.

What to watch next

Look for the follow-up experiments Google lists: larger codes, leakage removal, and tests that consume the cultivated states in an actual logical operation. Also watch whether rivals using other qubit types, such as trapped ions or neutral atoms, report their own cultivation or distillation results, since each platform has different noise. The number to track is the one that matters in the end: how many physical qubits a useful computation needs.

Sources

• Google Quantum AI: Bringing the magic (blog post, 15 Oct 2026)
• Science: Magic state cultivation on a superconducting quantum processor (DOI 10.1126/science.aee7961), cited via the coverage below; we did not open the paper itself
• The Quantum Insider, 16 Oct 2026
• Quantum Computing Report, 15 Oct 2026