Google Quantum AI researchers have implemented a resource-state preparation technique on a superconducting quantum processor. Their paper on magic state cultivation reports a 40-fold reduction in error and a measured state fidelity of 0.9999(1).
The result addresses a costly part of fault-tolerant quantum computing, but it is not a demonstration of a useful general-purpose quantum computer. The experiment retained 8% of its attempts, an important qualification alongside the fidelity figure.
Why magic states matter
Error-corrected quantum computers need a universal set of operations. The paper says the required non-Clifford gates carry a large resource cost in many leading error-correction designs.
Magic states provide the resources for those operations, but preparing sufficiently clean states has traditionally required resource-intensive distillation. Cultivation is designed as a more efficient alternative.
The Google-led team implemented cultivation on physical hardware, including a switch into a surface code. The researchers also developed a fault-tolerant measurement protocol to place a bound on the resulting state's fidelity.
The 40-fold reduction has a tradeoff
The paper reports that cultivation reduced error by a factor of 40. It measured state fidelity at 0.9999(1), while retaining 8% of attempts.
Google CEO Sundar Pichai described that result as 99.99% logical fidelity in his announcement and said the work was published in Science on October 15. The attached research preprint, first submitted in December 2025, lays out the experiment and its measurements.
The 8% retention figure matters because fidelity and yield answer different questions. The result shows the quality of the states that survived the protocol's checks, not that 99.99% of all runs succeeded.
The authors say the experiment establishes magic state cultivation as a viable approach to one of quantum computing's major engineering problems. It still represents progress on a component of fault-tolerant computation, not evidence that the full stack is ready for practical workloads.
