Google Quantum AI has shown on real hardware that "magic state cultivation" works. In a paper published today in Science, its Willow processor produced T-states with a record logical fidelity of 99.99 percent and cut errors about 40-fold compared with uncultivated states, according to Google's announcement.
Posts circulating on X turn that into "Willow slashes quantum resource costs 40x." That is the part to correct. The measured 40x is an error reduction in one state-preparation step. Google itself says that larger-scale tests are still needed before the saving in machine size is demonstrated.
Quick reference
| Question | Answer |
|---|---|
| What was shown? | Magic state cultivation running on Willow |
| Headline numbers | About 40x fewer errors; 99.99% logical fidelity for T-states |
| Qubits used | 54 of Willow's 105, per Scientific American |
| Published | Science, October 15, 2026; lead author Emma Rosenfeld |
| Viral claim | "40x cheaper" resource costs: not measured |
| Catch | 92% of attempts discarded, per an outside expert quoted by Scientific American |
| Still needed | Larger-scale demonstration; far lower error rates for real applications |
Why magic states are a bottleneck
A sphere with two faint overlapping outlines, picturing the superposition a magic state prepares for magic state cultivation
Quantum computers need a gate that classical machines cannot imitate. Scientific American quotes Rosenfeld: Clifford gates can be simulated by an ordinary computer, but "there is one particular gate that unlocks the quantum computer's power," a non-Clifford gate, most often the T gate. That gate is powered by a magic state, which is consumed when used and must be prepared very precisely, since any flaw in the state becomes a flaw in the gate.
The standard recipe, distillation, turns many imperfect states into fewer accurate ones. Google says up to 95 percent of a quantum computer's capacity might need to go to making them. That is the resource bottleneck researchers have chipped at since 2004, and why the 40x headline spreads so easily.
What cultivation does
Google's team proposed cultivation in 2024 (Craig Gidney, Noah Shutty and Cody Jones). Instead of distilling many states, you inject a noisy state into a small patch of qubits and run coordinated checks that ask, in Scientific American's phrasing, "Are you the right magic state: yes or no?" without collapsing the state. Failed attempts are thrown away; states that pass are spread across more qubits for stronger protection. The state starts small, gets checked, then grows.
Per Scientific American, without cultivation about four attempts in 1,000 produced a faulty state; with it, errors fell roughly 40-fold. Google adds that it demonstrated grafting a state from a distance-3 color code into a distance-5 surface code, showing purified states can survive the transfer.
Google also describes a measurement problem that took months to find: odd data traced to electrical drift and temperature noise invalidating the team's usual benchmark. The fix was extra control and a new technique that uses cultivation to measure itself.
How to read "99.99 percent" and "40x"
Two numbers carry the headline, and they mean different things. The 99.99 percent logical fidelity says that a prepared T-state is correct about 9,999 times in 10,000, which matches Scientific American's figure of roughly one error in 10,000 attempts. The 40-fold figure is a ratio against the same preparation without the cultivation checks: Scientific American puts the baseline at about four faulty states in 1,000, so cultivation moves the error rate from roughly 4 in 1,000 toward 1 in 10,000, with the exact ratio depending on how the paper defines the comparison.
Neither number is a cost. A cost would be expressed in qubits, in time, or in the number of physical operations needed per useful state, compared against distillation under the same assumptions. The announcement gives a footprint for the experiment, 54 physical qubits of Willow's 105, but it does not say how much smaller a full machine becomes. That comparison is a projection from the 2024 theory paper, which suggested large savings on paper; this experiment tests that the underlying protocol works on hardware. Google says as much: "we still need to demonstrate magic state cultivation at a larger scale to make it realistic for practical quantum computing."
What post-selection means for the result
Cultivation works as a filter. Each attempt runs checks, and attempts that fail are discarded. Brown's observation that 92 percent of attempts were rejected is therefore central, not a footnote. A state that is correct 99.99 percent of the time among survivors can still be expensive per surviving state if most attempts are thrown away. Whether that is a good trade depends on how fast the hardware can repeat attempts and how many parallel patches a machine can run.
The comparison with distillation is the open point. Distillation consumes many noisy states to make one good one; cultivation consumes many attempts in a small area to make one good one. Which uses fewer qubit-seconds at scale is exactly what a viral "40x cheaper" skips over. The honest framing is that cultivation has moved from a promising proposal to a working experiment, and the cost comparison is the next thing to measure.
Where this sits in Willow's story
Willow is Google's 105-qubit superconducting chip. Earlier public results around it include a December 2024 error-correction result in which larger encoded grids had lower error rates, and a 2025 "Quantum Echoes" claim of verifiable advantage. Those showed that the chip can hold information and beat classical simulation on a benchmark. The magic-state work addresses a different gap: even with good error correction, a useful computer must also supply a steady stream of expensive non-Clifford gates, and Google calls this the first demonstration on its hardware of a practical solution to that bottleneck.
It also shows the less glamorous side of the field. Months of confusing data turned out to be a measurement artifact from drifting electrical equipment and temperature, and the team had to invent a new benchmarking technique to trust its own numbers. For readers judging quantum claims, that is a useful reminder that benchmark validity is as hard as the physics.
The targets that remain
Rosenfeld's own figures set the distance still to cover. Breaking RSA-2048 would need magic states failing once in 10 billion tries. Molecule simulation, which she calls more attainable, needs one error in a million to one in 100 million. The team is at about one in 10,000. That is a gap of two to six orders of magnitude depending on the application, which is why experts treat this as a step and not an arrival. It also explains why the near-term pitch from Google is scientific applications, not cryptography.
Claimed versus verified
- 40x fewer errors, 99.99% fidelity: reported by Google and Scientific American from the Science paper. We have not read the paper itself.
- "40x cheaper" resources: not what the announcement says. It says the result brings practical quantum computing closer and that scale-up demonstrations remain.
- "Solves" the bottleneck: Google calls it the first demonstration of a practical solution; Brown says the technique works but discards 92 percent of attempts, which is a cost.
- Ready for useful work: no. Rosenfeld says breaking RSA-2048 would need magic states failing once in 10 billion tries, and simulating molecules one in a million to one in 100 million. The team is at roughly one in 10,000.
What it means
For security planning, nothing changes this week: the gap to cryptographically relevant machines is still orders of magnitude on this one metric. See our coverage of Trump's quantum executive orders and the 2030 encryption deadline, and for how factoring claims should be checked, RSA-260 factored: verify it yourself and the Claude RSA-896 claim. For other quantum infrastructure news, see the wireless quantum network over 21 km.
For researchers, it is evidence that an idea from theory survives contact with noisy hardware, with the usual caveat that the next step is scale.
Why the "40x cheaper" version spread
The aggregator version of this story says Willow "produces critical resources known as magic states at 99.99% fidelity while reducing their production costs by roughly 40x, according to company data." It is easy to see how that happens: Google's own post frames the bottleneck as a resource cost, then reports a 40-fold figure, and a summarizer joins the two. The sentence reads naturally and is wrong in a way that matters, because the 40-fold figure is an error ratio. If you are writing about quantum results, the safest habit is to quote the unit of every number: errors, qubits, time or dollars.
Another habit is to separate three kinds of claim: what the experiment measured (error rates on 54 qubits), what the authors infer from theory (cultivation should shrink the machine needed), and what remains to be shown (a larger-scale run with a real cost comparison). Google's post is careful to keep these apart, and its closing paragraph says plenty of technical challenges remain.
How to check it yourself
- Read the Science paper and compare the logical error rate with and without cultivation.
- Look for the post-selection (discard) rate, and whether the 40x is conditional on it.
- Check whether any figure compares total qubit-time or footprint against distillation.
- Treat resource-cost claims as unproven until such a comparison appears.
Related reading
- Trump's quantum executive orders
- RSA-260 factored: verify it yourself
- The Claude RSA-896 claim
- Wireless quantum network over 21 km
- AlphaEvolve and Gemini
Details are accurate as of October 15, 2026, based on Google's blog and Scientific American's report on the Science paper.