Google Quantum AI and collaborators report that measurement contextuality on the company’s Willow processor let them beat classical success limits in several bounded-resource games, including a hidden-linear-function test on grids “of up to 105 qubits.” The figures are Google’s. They appear in a Science paper published on 19 November 2026, “Quantum-classical separation in bounded-resource tasks arising from measurement contextuality,” and in a Google Research blog of the same day. They are not a claim of useful or commercial quantum advantage.
The paper says “we show that quantum contextuality enables certain tasks to be performed with success probabilities beyond classical limits.” Contextuality, as the blog puts it, is the idea that a measurement’s outcome can depend on which other compatible measurements are made with it. The team ran four related tests on Willow: the magic square game, a Bell–Kochen–Specker inequality, an N-player GHZ game, and a two-dimensional hidden linear function.
In the magic square game, Google says, the qubits won “98.30%” of the rounds (0.9830 ± 0.0001) across 100,000 plays of each of the nine versions, against a classical maximum of “88.89%.” The blog’s account is that any pre-written table can win at most eight rounds in nine, and that shared entanglement is what lets the players beat that bound, with noise still causing about one loss in sixty on this hardware.
Google also reports a single contextuality number, χ, from a Bell–Kochen–Specker inequality. In the 19 November blog, the team says it measured χ = 5.590 ± 0.008, past the classical bound of 4 and short of the quantum limit of 6. That figure is Google’s, from the blog that accompanied the paper.
The larger tests are where the 105-qubit claim sits. Google says it played the GHZ game with up to 71 players, one per qubit, staying above the classical bound as the number of players grew, and that it verified genuine entanglement across 45 qubits. It also says it ran the hidden-linear-function algorithm, a shallow-circuit task, on grids “of up to 105 qubits,” with success probabilities above one-half. Those are hardware results on structured games. They do not show that Willow has solved a useful computational problem faster than a classical machine.
The paper adds a benchmarking argument. “Our work introduces benchmarking of quantum processors using contextuality-based algorithms.” The blog says the same tasks “surfaced subtle coherent errors” that random-circuit benchmarks “tend to average away.” That is Google’s case for why these games may be a sharper ruler than a random-circuit score. It is not, on its own, evidence that the processor is ready for chemistry, materials or other applications the blog mentions as a longer-term interest.
The work sits next to earlier hardware claims that also have to be read inside their fence. Google Quantum AI has already reported “magic state cultivation” working on Willow, with a roughly 40-fold error cut in that experiment. The new blog itself points back to magic states as a theoretical link to contextuality. A different team’s Quantinuum result of 99.9995% logical fidelity on its Helios machine is another vendor-reported hardware figure, not a demonstration of useful advantage. So is NVIDIA’s simulation of 104 qubits on JUPITER to test a quantum error fix, which is a classical simulation of a correction, not a 105-qubit processor doing useful work.
A bounded-resource win is not useful quantum advantage
The sharpest sentences in the paper and the blog are also the ones that have to stay inside their setting. Google says Willow won the magic square game 98.30% of the time against an 88.89% classical cap, that χ came in at 5.590 ± 0.008, that a GHZ game stayed above the classical bound up to 71 players, and that a hidden-linear-function test ran on grids of up to 105 qubits with success probabilities above one-half. Those are bounded-resource tasks, designed so that a classical strategy with a pre-written table, or a shallow classical circuit, has a known ceiling. Beating that ceiling is a real hardware result if the data hold. It is not the same as showing that a quantum computer has done a useful job more cheaply, more quickly, or at a scale that matters outside the laboratory.
The blog is careful, in places, to describe noise and the narrowing lead as more qubits are added. That caution is the right one to keep. A contextuality score close to the quantum limit of 6, and a 105-qubit shallow-circuit run that stays above one-half, can still be a useful benchmark for the machine. They do not, without a named application, a comparison against the best classical method for that application, and an independent check, amount to useful quantum advantage. The next facts to watch are an independent reproduction of the published χ and win rates, or a result that takes the same hardware beyond a bounded game and into a task whose value does not depend on the game’s own scoring rule. Neither appears in the paper or the blog.
How we sourced this
The quoted findings and all performance figures are taken from the Science abstract and from Google’s 19 November blog, as set out in Research’s brief. χ = 5.590 ± 0.008 is the figure in that blog. No other χ value has been used.
Sources
- Google Quantum AI and Collaborators, “Quantum-classical separation in bounded-resource tasks arising from measurement contextuality,” Science, Vol. 394, Issue 6826, 19 November 2026.
- Shashwat Kumar and Rodrigo Cortiñas, “A game you cannot win unless you have a quantum computer,” Google Research, 19 November 2026.
Get TSN’s biggest tech stories, explained.
News and insights across AI, tech, crypto and cybersecurity. Every source linked. Unsubscribe anytime.
Get the newsletterRelated stories
Google Quantum AI Shows “Magic State Cultivation” Working on Its Willow Chip, Cutting Errors About 40-Fold is the closest recent TSN account of a Google Quantum AI hardware result on Willow, and is the next read.
- Quantinuum Reports 99.9995% Logical Fidelity on Its Helios Machine
- NVIDIA Simulates 104 Qubits on JUPITER to Test a Quantum Error Fix
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X

