Google's Willow Processor Wins the Magic Square Game 98.3% of the Time: What the Science Paper of 19 November 2026 Shows About Quantum Contextuality at 105 Qubits

Board-ready intelligence on quantum innovation · Biomedical discovery · Post-quantum transition
Google Quantum AI's Science paper of 19 November 2026 measures quantum contextuality on Willow, from a two-player game to 105-qubit shallow circuits. This post separates what the experiments establish from what the announcement leaves open.

Quantum Governance

Google Quantum AI's Science paper of 19 November 2026 measures quantum contextuality on Willow, from a two-player game to 105-qubit shallow circuits. This post separates what the experiments establish from what the announcement leaves open.

Published by Quentir Systems LLC · November 21, 2026 · 7 min read

In 2022 the Nobel Committee awarded the physics prize to Alain Aspect, John Clauser and Anton Zeilinger for experiments with entangled photons that ruled out local hidden-variable explanations, in which particles carry their measurement results with them from the start, under the assumptions those experiments were built to test. They gave a sharp experimental answer, for local theories, to a question Einstein had pressed for decades: whether physical properties exist before anyone measures them. Four years later a closely related test has moved from the foundations laboratory into the engineering of quantum computers. On 19 November 2026 Google Quantum AI published in Science a set of experiments that use quantum contextuality, a broader property of which the nonlocality in Bell tests is a special case, as a working measure of what its processors can do.

Practical takeaway. Google's Willow processor beat the best classical strategy in two games played without communication, with up to 71 players, and stayed above a conjectured classical bound in a shallow-circuit task on up to 105 qubits. Every comparison is with classical strategies under the same resource limits. The result establishes that a nonclassical resource survives at scale on noisy hardware and gives a benchmark whose target is fixed by physics. It establishes no speed or cost advantage on a practical problem, and it does not move any post-quantum migration deadline.

What Google Quantum AI reported in Science on 19 November 2026

The paper, "Quantum-classical separation in bounded-resource tasks arising from measurement contextuality", appears in Science volume 394, issue 6826, with a large Google Quantum AI author list and academic collaborators. Research scientists Shashwat Kumar and Rodrigo Cortiñas summarized it the same day on the Google Research blog. Contextuality is the experimentally testable statement that a quantum system cannot have fixed, pre-existing values for everything one might measure that do not depend on the measurement context: the result of one measurement can depend on which other compatible measurements are made alongside it. Theory has linked this property to the advantage some quantum algorithms hold over classical ones, but until recently it had been shown almost only in systems of a few particles. The Google team asked whether it survives in the many-qubit devices on which quantum computing depends.

The magic square game on Willow: 98.30 percent against a classical ceiling of 88.89 percent

The first experiment uses a puzzle described by Asher Peres and David Mermin around 1990. A three-by-three grid must be filled with +1 and −1 so that every row multiplies to +1 and every column to −1, which is arithmetically impossible. In the game version, one player receives a row and another a column, each fills in entries without communicating, and they win when they agree on the shared cell. Any strategy written in advance is a fixed table and loses at least one round in nine, so the classical ceiling is 88.89 percent. Two players each holding two qubits from shared entangled pairs can in principle win every round. On Willow, all nine versions of the game were played 100,000 times each, and the qubits won 98.30 percent of rounds, losing about one in sixty because of hardware noise.

The team then quantified contextuality with a Bell–Kochen–Specker inequality that combines the row and column products into one number, χ. Any noncontextual assignment of predetermined outcomes keeps χ at or below 4, and quantum mechanics allows up to 6. Measuring it required every row and column to be read out on the same two qubits in sequence, using a helper qubit so the main pair was left undisturbed, over 540 consecutive measurements repeated 1,000 times. The measured value was χ = 5.590 ± 0.008, about 7 percent below the quantum limit. Because the test is state-independent, the score describes the machine rather than a carefully prepared input.

From two qubits to 105: the GHZ parity game and the hidden linear function problem

The second family tests scale. In the GHZ parity game, every player holds one qubit of a shared entangled state and answers a one-bit question with one bit. The best classical plan wins 75 percent of the time with three players and approaches 50 percent as players are added. Google played with up to 71 players and stayed above the classical bound throughout, although noise narrowed the margin with every added qubit; the team also verified genuine entanglement across 45 qubits. The third family runs the problem from a 2018 theorem by Sergey Bravyi, David Gosset and Robert König, which proved that quantum circuits of constant depth solve a hidden linear function problem that classical circuits of constant depth, built from gates with bounded fan-in, cannot solve with high probability. The theorem is asymptotic; any experiment at a fixed size is compared with a finite-size bound. Google ran that algorithm on grids of up to 105 qubits. Success probability fell from close to 1 to about 0.6 at full processor size, which the team reports as still above the conjectured classical bound.

Which claims the paper supports and which it leaves to later work

Read against Google's own account, the paper supports three claims. Contextuality, a nonclassical resource, is present and measurable on processors of up to 105 qubits. In the two games the separations are statistically significant against the best classical strategies without communication between players; in the hidden-linear-function task the comparison is with a conjectured classical bound for constant-depth circuits at that size. The games also proved to be sensitive diagnostics: they surfaced small coherent errors, rotations of the quantum state slightly off target, that random-circuit benchmarks tend to average away.

Several further statements belong to expectation. The separation says nothing about how fast or how cheaply Willow solves a problem anyone needs solved, and the blog's title, a game "you cannot win unless you have a quantum computer", describes a game built to have exactly that property. The classical bound for the 105-qubit task is conjectured, so a later classical result could narrow the claimed gap, although the theorem for constant-depth circuits stands. Google's view that the benchmark is relevant to chemistry and materials simulation, and the theoretical link between contextuality and the magic states that error-corrected machines need, are reasonable research directions that this paper does not test. The Science text sits behind a paywall; this post maps its figures to the authors' public summary and to the journal's bibliographic data.

How Quentir Reads It

The most useful product of the paper for anyone outside physics is the measure itself. Most quantum benchmarks ask whether a device clears a classical bar, and classical bars keep moving as classical algorithms and supercomputers improve. The contextuality games give a ceiling set by physics: a win probability of 1, or χ = 6, and the distance between a machine's score and that ceiling stays meaningful as classical computing improves. Scores from different machines are comparable only when the protocol, the number of rounds, the error bars and any discarded runs are fixed and reported in advance. For a public research funder, a national laboratory buying a system, or an investor comparing announcements, a score reported under those conditions says more about a machine's quality than a headline qubit count. The games and their classical bounds are textbook results, so any group with comparable hardware can run them and compare scores. That stands apart from the Google cryptanalysis estimate that Quentir examined in July, which was published as a public claim built on a private attack circuit and backed by a zero-knowledge proof.

The result also fits a sequence. In October Google reported in Science that magic-state cultivation worked on Willow at small scale, a hardware test of an assumption inside the estimate that fewer than one million noisy qubits could break RSA-2048. The contextuality paper measures a different resource and moves no cryptographic estimate. Post-quantum migration plans should therefore keep the schedules that apply to each organization. In the United Kingdom the effective practices that the Bank of England, PRA and FCA published for financial firms on 19 November are non-binding and discuss firms' migration plans without setting a deadline of their own. The case for migrating now rests on data harvested today and decrypted later, on signatures and authentication that must stay trustworthy for years, and on the long lead times of replacing cryptography across products and suppliers. Each laboratory milestone is a reason to check that cryptographic systems can be changed without rebuilding them. Signature Brief 2026.4, Crypto-Agility as a Verifiable Duty, sets out how that ability can be verified, through maturity models with published evaluations, quantified agility metrics and the gaps in existing supplier contracts.

Sources: Google Quantum AI and collaborators, "Quantum-classical separation in bounded-resource tasks arising from measurement contextuality", Science 394 (6826), published 19 November 2026 (bibliographic data via Crossref; figures as reported in the authors' summary); Shashwat Kumar and Rodrigo Cortiñas, "A game you cannot win unless you have a quantum computer", Google Research blog, 19 November 2026; Sergey Bravyi, David Gosset and Robert König, "Quantum advantage with shallow circuits", Science 362 (6412), 2018; The Nobel Prize in Physics 2022, NobelPrize.org. Links checked 21 November 2026.

Published intelligence, built to inform your own decisions. Published: November 21, 2026.

Next
Next

GPT-4o, Claude Sonnet 5 and Qwen3.7-plus Defer to Senior Clinicians in an Oxford and Stanford Study: What It Means for Human Oversight Under AI Act Article 14