Google's 105-Qubit Quantum Processors Win Contextuality Game at 98.3%
Google Quantum AI won a contextuality game 98.3% of the time across 105 qubits, above the 88.9% classical limit, in a Science cover study that also exposes coherent hardware errors.

Why it matters
- Google's magic-square game won 98.3% of the time against an 88.9% classical limit, published on the cover of Science and announced November 19, 2026.
- Contextuality tests extended to as many as 105 qubits; prior demonstrations mostly involved only a few particles.
- A Kochen–Specker–Bell test measured χKSB = 5.618 ± 0.005, above the classical bound of 4 and near the quantum limit of 6.
- GHZ parity games ran through 71 qubits, with the quantum-classical margin narrowing as players were added.
- Effective depth on the hidden-linear-function task rose from an ideal four gate layers to about six on the largest Willow grid.
- The classical lower bound for hidden-linear-function becomes relevant only at roughly one million qubits.
Google Quantum AI has demonstrated measurement contextuality on as many as 105 qubits, winning a coordination game 98.3% of the time on its superconducting processors — a result published on the cover of Science and announced on November 19, 2026. The score clears the 88.9% ceiling that any classical strategy can reach, and the experiments behind it stretch a property of quantum mechanics that had previously been confirmed mostly in systems of just a few particles.
Contextuality means a measurement outcome is not a predetermined value; it depends on which other compatible measurements are performed alongside it. Theory suggests this property may power certain quantum algorithms, which is why demonstrating it at scale matters for both research and the commercial race to build useful quantum machines. The underlying work first appeared as an arXiv preprint on December 1, 2025; the November 2026 announcement marks its journal publication rather than the initial disclosure of the experiments.
What is the magic-square game?
The experiments open with the Mermin–Peres magic-square game. Two players, Alice and Bob, agree on a strategy before being separated. A referee then assigns Alice a row and Bob a column of a three-by-three table. Alice's entries must multiply to +1 and Bob's to −1, and the players win when their values match at the intersecting cell.
No prearranged classical table can satisfy those opposing constraints for every assignment. That is the 88.9% classical limit. The quantum implementation uses a two-by-two square of qubits and two shared Bell pairs (entangled qubit pairs shared between the players), with an ideal noiseless strategy winning every game. The reported experimental average came from nine row-column assignments, each repeated 100,000 times.
A separate Kochen–Specker–Bell test quantified contextuality with χKSB, a combination of correlations from compatible measurements. The paper reports χKSB = 5.618 ± 0.005, exceeding the classical bound of 4 and approaching the quantum limit of 6. The authors attribute the remaining gap to experimental errors. Its circuit streamed 180 randomly selected measurement contexts over 1,000 runs.
How did the tests scale?
To push contextuality beyond few-qubit states, the team implemented an N-player GHZ parity game. Each player corresponded to one qubit, and the group shared a GHZ state — a multi-qubit entangled state — across those qubits. Each player received a binary question and returned a bit. To win, the parity of their combined answers had to match a value determined by the questions.
The paper reports tests through 71 qubits, with the quantum winning probability above the classical probability at every tested size. The advantage nevertheless narrowed as players were added. For each player count, the researchers tested 20 games across 50 qubit configurations, repeating each configuration 1,000 times to distinguish real performance from statistical fluctuations as the margin contracted.
A separate two-dimensional hidden-linear-function task asked the processor to find a hidden bitstring associated with a symmetric matrix. The experiments used Willow, Google's superconducting processor family, and controlled-Z gates as the native entangling operation, testing sizes from 9 qubits up to the largest device. At each size, the team measured 1,000 random matrices and ran each matrix 100 times.
The results quantify the cost of scale. An error-free implementation returns a correct bitstring after four effective two-qubit-gate layers. The researchers converted the fraction of correct outputs into an effective depth — a time-to-solution measure reflecting the extra runs needed to obtain a correct answer. On the largest grid, that measure rose to about six layers.
Why do the tests double as diagnostics?
The authors propose contextuality-based algorithms as processor benchmarks, and the many-body versions exposed subtle coherent errors that random unitary operations (randomly applied quantum logic gates) did not reveal. That gives the tests a hardware-diagnostic role beyond their game scores: a processor can pass unstructured checks while failing ones rooted in the very property thought to drive quantum speedups.
Google presents that error-detection capability as a step toward future chemistry and materials-science applications. Those are stated potential uses. The reported experiments do not simulate either application.
What are the limits of the quantum-classical separation?
The scope of the claimed separation is equally important. The paper states that no unconditional proof of quantum advantage exists for a general computational model; its results concern particular tasks with bounded resources.
For the hidden-linear-function comparison, the established classical lower bound becomes relevant only at roughly one million qubits — far beyond current devices. The logarithmic near-term classical-depth bound shown against the experimental results is expressly a conjecture. The measurements therefore do not establish an unconditional advantage over every classical implementation at current processor sizes.
What happens next?
The immediate value is a sharper way to evaluate quantum hardware. Contextuality supplies scalable tasks tied to a proposed source of computational advantage, and their sensitivity exposes errors that can remain hidden in less structured tests. At the same time, the shrinking GHZ-game margin and the rising effective depth — from four layers in the ideal case to roughly six on the largest grid — show that preserving the quantum benefit becomes harder as more qubits participate. Closing that gap, rather than publishing another winning score, is the benchmark that will determine whether contextuality-driven algorithms move from Science covers into working applications.
Source: unite.ai
More from Marcus Bennett
Show full bio
Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
791 articles
Related articles
- MIT Plugs GPT-5.6 Sol Into Quantum Lab, and It Calibrates Qubits Overnight
- Zapata and UMD Launch ForShor: AI Quantum Designs Must Pass Lean Proofs
- AI Research Agents Reach Just 15% of Human Benchmark, Study Finds
- Moth’s Tessa encodes 1,568 values on a 120-qubit chip
- Ataraxos Beats Stratego's Greatest Player, Ending a Human Stronghold