We may earn a commission when you buy through links on this page. It doesn't change what we recommend.
Quantum computing headlines tend to arrive with a big multiplier attached: 40 times fewer errors, 12 times fewer gates, a calculation that would take a supercomputer longer than the age of the universe. The numbers are usually real, but they measure narrow things. This explainer covers the vocabulary behind them, error correction, logical qubits, gate counts and magic states, and what two October 2026 results do and don’t prove.
In this explainer
- Why quantum computers need error correction
- What “below threshold” means
- What gate counts measure
- Magic states: the most expensive ingredient
- What the October 2026 results show
- What they don’t prove
- How to read the next quantum headline
Why quantum computers need error correction
A quantum computer stores information in qubits, which can be built from single atoms, superconducting circuits or particles of light. Qubits are extremely sensitive. In Anthropic’s description, heat, vibration and even cosmic rays can nudge a qubit’s state, and those small errors pile up until the computer gives the wrong answer. The machines built so far have only hundreds to thousands of qubits, the same post says, too few for many of the algorithms that promise a practical advantage.
Error correction is the answer the field is betting on. Instead of trusting one physical qubit, a machine spreads a single unit of information across a group of physical qubits and keeps checking the group for errors. That protected unit is called a logical qubit. The catch is overhead: every logical qubit costs many physical ones, plus the time spent checking and fixing.
What “below threshold” means
Adding qubits normally adds errors. In Google’s announcement of its Willow chip in December 2024, the company said the more qubits you use, the more errors typically occur. Error correction only pays off if bigger groups of qubits make the logical qubit better, not worse. Reaching that point is called being “below threshold.”
Google reported in Nature that Willow, a 105-qubit chip, crossed it. As the team scaled its encoded qubit from a 3×3 grid to 5×5 to 7×7, it said, the error rate was cut in half at each step. Google called this a challenge the field had pursued since error correction was introduced in 1995. It is a milestone for the method, not a useful computer: the same post says the next challenge is a first “useful, beyond-classical” computation relevant to a real-world application.
The famous Willow speed figure, a benchmark computation in under five minutes that Google estimated would take a leading supercomputer 10 septillion years, used a test called random circuit sampling. Google itself says that benchmark has no known real-world application yet.
What gate counts measure
A quantum program is a long sequence of simple operations, called gates, applied to qubits. Fewer gates means a shorter run and fewer chances for errors. Fewer qubits means the program fits on a smaller machine. Researchers trade one against the other all the time.
Anthropic’s post gives a useful example of why gate counts matter. In 2017, researchers estimated that simulating FeMoco, an iron-sulfur cluster at the heart of the enzyme bacteria use to turn nitrogen into ammonia, would take about 50 trillion Toffoli gates, one of the most expensive gate types. Through years of algorithm work by many groups, that estimate has fallen to about 340 million, roughly a 150,000-fold reduction, which the post says corresponds to a few hours on a future machine with about three million physical qubits. No such machine exists yet. The reductions came mostly from redesigning the circuits, not from better hardware.
Two rules make quantum circuits harder to design than ordinary code. A quantum circuit can’t take a shortcut with an “if-then” branch, because it runs on many inputs at once, so it has to include the operations for every case. And it can’t simply delete its scratch work: every step must be reversible, so leftover information has to be “uncomputed” by running steps backward, which costs more gates. That is why one design can save qubits by spending gates, and another can do the reverse.
Magic states: the most expensive ingredient
Error correction protects most operations fairly cheaply. A few, including a type called the T gate, need special inputs called magic states. Google’s Quantum AI team says magic states are hard to make accurately, can be used only once, and that some estimates suggest up to 95% of a quantum computer’s capacity might be needed to produce them. Those are estimates, not measurements on a working machine.
The traditional fix, distillation, uses dedicated sections of the machine to turn many rough magic states into a few good ones. In 2024, Google researchers Craig Gidney, Noah Shutty and Cody Jones developed an alternative they call magic state cultivation, which works more like a filter: a rough state is placed in a small patch of the chip, coordinated measurements check it for errors, and failed attempts are thrown away.
What the October 2026 results show
Google: magic state cultivation on real hardware. In a paper published in Science on Oct. 15, Google reported that cultivation on Willow reduced errors about 40-fold compared with uncultivated states and produced T-states with a record logical fidelity of 99.99%, about one error in 10,000. The team also moved purified states from one error-correcting code (a distance-3 color code) into another (a distance-5 surface code), a step it calls grafting. Our news story covers what Google’s magic state cultivation result does and doesn’t prove, including the caveats reported alongside it.
Anthropic: AI-designed circuits. In a research post with a paper on arXiv, Anthropic said its Claude Mythos 5 model, working in a closed loop with an automated checker, designed circuits for eight basic routines that used fewer qubits and fewer gates than the best previously published designs. The most expensive step of an optimization algorithm called decoded quantum interferometry became about 12 times cheaper in gates than Google’s previous construction. One elliptic-curve design uses 12% more qubits than the smallest earlier design but 17 times fewer gates. Our news story explains what the 12x gate cut does and doesn’t mean.
Both results attack overhead rather than raw hardware. One makes a costly ingredient cheaper to produce; the other makes the recipe use less of everything.
What they don’t prove
- Neither is a useful quantum computer. Google says cultivation still has to be demonstrated at larger scale and that “plenty of technical challenges remain.” Anthropic says circuit design is only one part of the problem, alongside error-correcting codes, fast decoders and hardware control.
- Both are self-reported. Google’s work appeared in a peer-reviewed journal. Anthropic’s post does not say whether its paper has been peer reviewed. In both cases, replication by other groups is the real test.
- The multipliers are task-specific. Google’s 40-fold figure applies to preparing magic states, not to Willow’s overall performance. Anthropic’s 12x figure applies to one decoder step, and its score, qubits multiplied by gates, is one reasonable choice among several.
- Encryption is not broken. The elliptic-curve circuits matter because point addition is the core step in quantum attacks on elliptic-curve cryptography, but a better circuit is still only a design. It needs a large error-corrected machine to run on, and none exists.
- Timelines are expectations. Anthropic says useful quantum computers “could arrive sooner than previously anticipated.” Google gives no date. Neither is a demonstrated outcome.
How to read the next quantum headline
Five questions sort most announcements quickly:
- Is the number about physical qubits or logical qubits?
- Was it measured on hardware, or estimated in simulation?
- Which resource shrank: qubits, gates, run time or errors, and on which task?
- Has anyone outside the company reproduced it, and has it been peer reviewed?
- Does the company say what is still missing? The credible ones usually do.
Progress in this field usually looks like the two October results: a specific, measured improvement on a specific bottleneck, with the caveats printed in the paper. That is real progress, and it is not the finish line.
