Illustrative image: a quantum circuit diagram — qubits (horizontal lines) passing through unitary gates. The circuits in this story are fault-tolerant designs of this kind, not photographs of hardware. Image: Wikimedia Commons, CC BY-SA 4.0
📌 KEY HIGHLIGHTS
- New on 10 November 2026: QuEra Computing highlighted the joint Harvard–Anthropic–QuEra research in a press release and blog post; the preprint itself was posted to arXiv on 15 October and Anthropic's technical account followed on 16 October.
- Seven of eight tasks improved: the AI's circuits use fewer Toffoli gates AND fewer qubits than the best previously published designs.
- Biggest gains against published designs: 4.1x to 39x lower space-time cost; one decoder also scored more than 9,100x lower — but only against a baseline the authors built themselves, not a published design.
- 53-hour run: a single autonomous run produced a new reversible design for elliptic-curve arithmetic — 17 times fewer gates than the smallest previous design.
- Honest caveat: these are circuit designs verified by classical simulation — nothing in the paper was run on a quantum computer.
What happened?
On 10 November 2026, QuEra Computing highlighted joint research with Harvard University and Anthropic showing that an AI can design quantum circuits that need fewer resources than the best published human designs. The work itself was posted as a preprint, "Discovering state-of-the-art quantum circuits with large language models", to arXiv on 15 October 2026, with Anthropic publishing its technical account the next day. The team — Andi Gu (Harvard, in Anthropic's STEM Fellows Program), Anthropic quantum physicist Alek Kemeny, Susanne Yelin of Harvard and Mikhail Lukin of Harvard and QuEra — set out to test whether an AI could do one of the most expert-heavy jobs in quantum computing: designing the circuits that a future error-corrected machine would actually run.
The model doing the designing was Claude Mythos 5, working inside an open-source evolutionary search framework called OpenEvolve. The eight tasks it tackled are the basic building blocks of larger quantum algorithms, drawn from chemistry, optimisation, physics simulation, number theory and numerical computing. On seven of them it beat the best published human designs; on the eighth it invented a new architecture altogether.
QuEra Claude quantum circuits explained: how the AI designs circuits
Think of a quantum circuit as a recipe: qubits are the ingredients, and gates are the steps. For a future fault-tolerant machine, the expensive steps are Toffoli gates — a particular three-qubit operation that is costly to run error-free — and the number of qubits the recipe needs at once. The researchers score every candidate circuit by its "TQ cost": Toffoli gates multiplied by peak qubits. Lower is cheaper, in both time and hardware.
The loop works like this. Researchers hand Claude a plain-language specification of the circuit they want. Claude writes a candidate design as a Python program. An independent evaluator — a classical program, not the AI — checks whether the design is correct and scores its TQ cost. The best designs survive into the next round, and Claude revises from there, with no human input between rounds. Good ideas are also shared across competing design families, so a trick found for one task can migrate to another.
That last detail is what makes the result more than a faster search. In the DQI decoder task — a step in a leading quantum optimisation algorithm — Claude ran 403 million sample decodings, discovered that almost every run needed only one saved value, and reserved 19 qubits instead of a worst-case 665. It also swapped the decoding method from the extended Euclidean algorithm to Berlekamp–Massey decoding and rebuilt temporary values from the circuit's current state instead of storing them. The step became about 12 times cheaper than Google Quantum AI's previous construction, with about one run in 10,000 failing — inside the task's allowed error budget.
| Task | What it is | Reported space-time cost gain |
|---|---|---|
| Hermitian DQI decoder | Decoder for a quantum optimisation algorithm | More than 9,100x lower — vs the authors' own baseline, not a published design |
| Fixed-point Gaussian | Numerical function used inside quantum algorithms | 39x lower |
| HHL rotation angle (arcsin) | Step in a quantum linear-equations solver | 32x lower |
| Reed–Solomon DQI decoder | Decoder for the same optimisation algorithm family | 12x lower |
| Coulomb potential | Chemistry simulation building block | 9.9x lower |
| Non-Abelian lattice-gauge simulation | Physics simulation of fundamental forces | 7.7x lower |
| Fe2S2 alias-select | Chemistry step for an iron-sulfur molecule | 4.1x lower |
| Prime-field modular inversion | Arithmetic step in elliptic-curve cryptography | New architecture (no single baseline gain) |
Illustrative image: a quantum computer at Chalmers University of Technology (2017). The circuits in this story are designs, not runs on this or any other machine. Image: Wikimedia Commons, CC BY-SA 4.0
The 53-hour run that rewrote elliptic-curve arithmetic
The eighth task is the one cryptographers will notice. Modular inversion is the costliest step of elliptic-curve point addition — the arithmetic behind much of today's public-key cryptography. Claude worked on it for a single 53-hour run with minimal human input and produced what Anthropic calls the first reversible implementation of divsteps with no stored branch history, checked by simulation and formalised in the Lean proof system. One complete benchmark circuit in the resulting family uses about 4.0 million Toffoli gates and 935 qubits; one design uses 12% more qubits but 17 times fewer gates than the smallest previous design.
That does not mean your encrypted messages are at risk. These are cheaper blueprints for arithmetic a fault-tolerant quantum computer would one day need — and no such machine exists yet. But it does move the timeline question: every order-of-magnitude cut in circuit cost is a cut in the hardware that a future code-breaking machine would need. Related work in the same direction includes Iceberg Quantum's Pinnacle 1.2 decoder cutting RSA-2048 factoring resources and NVIDIA's JUPITER 104-qubit quantum simulation — the whole field is converging on cheaper fault tolerance.
Why it matters for World Tech readers
Useful quantum computing has always needed two things: better hardware and smarter circuits. Hardware gets the headlines; circuits have quietly done much of the heavy lifting. Anthropic's favourite example is FeMoco, the molecule bacteria use to turn nitrogen from the air into ammonia. A 2017 estimate put simulating it at about 50 trillion Toffoli gates. Eight years of human algorithmic work brought that to about 340 million — roughly 150,000 times fewer — without any change to the hardware. This paper suggests AI can now do that kind of work, and quickly: what took expert teams months per step took Claude hours to days per task.
For readers in India and elsewhere, the practical angle is the same as for every quantum advance: the payoff is years away, but the cost curve is what decides when it arrives. Cheaper circuits mean smaller, sooner machines for drug discovery, materials design and optimisation — and, eventually, a harder conversation about upgrading today's encryption. Anthropic is candid that Claude started conservative and needed prompting to explore genuinely new approaches, and that these results are designs awaiting hardware, not demonstrations. That honesty is part of why the paper is worth taking seriously. It is also not the only AI-designed quantum result this week: Google DeepMind reported AI agents finding quantum programs up to 290x cheaper, with correctness proved in Lean — a different method, the same direction of travel.
Should you care?
FAQs
What is a quantum circuit?
A quantum circuit is a step-by-step recipe for a quantum computation: horizontal lines represent qubits, and boxes on those lines represent gates — operations that change the qubits' states. The circuit's width is its qubit count and its depth is its number of gate layers.
What is a Toffoli gate, and why count them?
A Toffoli gate is a three-qubit operation that flips a target qubit only when two control qubits are both set. On an error-corrected quantum computer it is one of the most expensive operations to perform, so Toffoli count — multiplied by peak qubits — is the standard proxy for how costly a circuit will be to run.
Did Claude run these circuits on a real quantum computer?
No. The designs were checked by classical simulation and, for the divsteps result, formalised in the Lean proof system. They are verified blueprints, not hardware demonstrations — an important distinction the paper itself maintains.
Does this mean quantum computers will break encryption soon?
No. Cheaper elliptic-curve arithmetic lowers the eventual hardware bar, but a fault-tolerant machine large enough to run these circuits does not exist yet. The result shortens the road; it does not arrive at the destination.
Takeaway: AI just beat the best human quantum-circuit designs on seven of eight tasks, with gains of 4.1x to 39x against published designs — and it did the work in hours and days, not months. The designs still await real hardware, but the cost curve of useful quantum computing just bent a little steeper.
Want the full numbers? The preprint's tables list every task — read it on arXiv — and for the week's other AI-designed quantum result, see our Google DeepMind Lean explainer.
Sources
- Gu, Kemeny, Yelin & Lukin — "Discovering state-of-the-art quantum circuits with large language models" (arXiv, 15 Oct 2026, primary source)
- QuEra Computing — "AI Is Cutting the Cost of Quantum Algorithms" (company blog on the joint work)
- Anthropic — "Designing more efficient quantum circuits with Claude" (16 Oct 2026, primary source)
- The Quantum Insider — "QuEra, Harvard and Anthropic Use AI to Reduce Quantum Circuit Resource Requirements" (10 Nov 2026, QuEra press release)
- Quantum Computing Report — "Harvard, Anthropic, and QuEra Use LLMs to Discover Quantum Circuits" (17 Oct 2026)
- Unite.AI — "Anthropic Says Claude Beat Best Published Quantum Circuit Designs" (Oct 2026)
Comments
Post a Comment