Google says it cut errors 40-fold in a quantum state-preparation experiment

Google Quantum AI says it has demonstrated a way to prepare crucial quantum states with about 40 times fewer errors on its Willow processor. In an October 15 research announcement, the team described turning “magic state cultivation” from a theoretical proposal into a hardware experiment aimed at reducing a major resource burden in quantum computing.

The problem starts before a quantum computer can do its useful work. To carry out the full set of operations needed for universal quantum computing, it needs special prepared states known as magic states. Creating high-quality versions protected by error correction consumes substantial resources. Each state can be used only once, so the machine must keep producing them rather than reuse a fixed supply.

Researchers have recognized that bottleneck since 2004. Google cites estimates suggesting that making magic states could occupy up to 95% of a quantum computer’s capacity, leaving little for the calculation itself. That is an estimate of the potential burden, not a measured capacity saving from the new Willow experiment.

In 2024, Google researchers Craig Gidney, Noah Shutty and Cody Jones refined an idea from earlier research into magic state cultivation. Their proposal treated preparation as a filtering process: start with an imperfect state, check it carefully, and retain attempts that pass. The intended payoff was a smaller quantum computer for practical tasks.

The cultivation sequence

1. Inject a noisy quantum state into a small region of the chip.
2. Run coordinated measurements to check for errors and discard flawed attempts.
3. Grow the low-error state while adapting it into a larger encoding—the form used to protect its quantum information.

Getting that sequence to work exposed a problem in how the team measured success. Research scientist Emma Rosenfeld initially saw data that appeared to defy physics. After months of simulations, discussions and checks, she traced it to an incorrect assumption in the team’s usual measurement technique. Noise from electrical-equipment drift and temperature changes was undermining the benchmark’s validity.

The response involved both hardware control and a different way of checking the result. The researchers added controls to reduce the noise’s effects and developed a benchmarking technique that used cultivation to measure itself. Google says those changes, alongside other engineering and algorithmic work, enabled the reported demonstration.

99.99%T-state logical fidelity

Google reports 99.99% logical fidelity for its prepared T-states, a measure of how closely those states match the intended state.

The experiment also addressed how a cultivated state could move between two forms of quantum error correction. The team transferred states from a distance-3 color code to a distance-5 surface code. Google describes color codes as well suited to cultivation and surface codes as a more established approach. This “grafting” step tested whether purified states could survive the transfer inside its hardware.

Google says the work was published in Science on October 15, 2026. It calls the result a breakthrough, but says larger-scale demonstrations are still needed to make cultivation practical. The experiment establishes the technique on hardware; the resource-reduction goal still requires further scaling, with technical challenges remaining.

For readers looking to inspect the underlying material, an associated data record on Zenodo lists files for Figures 2, 3 and 4 and an example-usage notebook. The record names Google Quantum AI as the creator and is dated December 16, 2025, predating this week’s announcement.