One step in a quantum chemistry calculation used to need 6,355 of the hardest quantum logic gates. A new design needs 1,140. That’s the kind of cut showing up across a batch of Claude quantum circuit designs described in a research preprint that Harvard, QuEra Computing and Anthropic researchers posted on arXiv in mid-October. The Quantum Insider covered the work again on November 10, which is what puts it in this week’s news.
The setup was simple. Give Claude, Anthropic’s AI model, a plain-language description of a circuit task. Pair it with an automated checker. Let it iterate on its own. In seven of eight circuit types tested, Claude’s designs needed fewer qubits and fewer of the costly logic gates called Toffoli gates than the best published designs. None of this ran on a real quantum computer. These are resource estimates from classical simulation and formal proofs, not hardware results, and the paper has not been peer reviewed.
How Claude’s Quantum Circuit Design Loop Worked
Strip away the jargon and the method breaks into a few clear steps:
- Researchers wrote the specs. Andi Gu led the work at Anthropic, with mentorship from physicist Alek Kemeny, in collaboration with Harvard’s Susanne Yelin and Mikhail Lukin, who holds appointments at both Harvard and QuEra. The team wrote plain-language specifications for eight circuit building blocks, covering chemistry, optimization, physics simulation and number theory.
- Claude proposed a design. For each task, Claude wrote a candidate circuit as a short Python program.
- An evaluator checked the work. A separate, automated program confirmed the design was correct for every input, then scored it by multiplying its Toffoli gate count by its peak qubit count, a rough stand-in for how long the circuit would take to run.
- Claude tried again. The score and any errors went back to Claude for another round, with no human stepping in between attempts.
- The hardest case got extra scrutiny. For a reversible approach to the modular inversion used in elliptic-curve calculations, Claude worked almost 53 hours straight with minimal human input. Researchers later verified the key step using the Lean proof assistant, software that checks mathematical reasoning line by line rather than just running it.
QuEra and Anthropic have worked together before. Earlier this year, QuEra used Claude-written software to fix laser faults in its own quantum hardware in seconds instead of minutes. This time, the collaboration moved from patching hardware problems to redesigning the software that runs on top of it.
The Resource Cuts, by the Numbers
Four of the results are worth comparing side by side. Each figure is a reduction in a combined cost metric, Toffoli gates multiplied by peak qubits, not a straightforward speed increase. How much faster any of this actually runs depends on hardware that doesn’t exist yet.
| Task | Resource-Cost Cut | Qubits, Before → After |
|---|---|---|
| Molecular simulation (Coulomb potential) | 9.9x fewer combined resources | 252 → 142 |
| Numerical computing (fixed-point Gaussian) | 39x fewer, the largest cut in the paper | 362 → 139 |
| Optimization decoding step (used in DQI) | About 12x fewer Toffoli gates | Not disclosed |
| Elliptic-curve point addition | 17x fewer gates, for 12% more qubits | 838 → 843 |
Source: Anthropic and QuEra Computing research posts, both Oct. 16, 2026. Figures are self-reported by the research team and have not been independently verified on hardware.
For elliptic-curve point addition, Claude actually produced two different designs, each trading gates for qubits in its own way. The table above shows the version built to match the lowest qubit count published so far. A second Claude design, based on a classical algorithm called divsteps, goes the other direction: it uses 35% fewer qubits than the design with the fewest published gates, at the cost of needing more gates overall. Nobody had made divsteps efficient enough for a reversible quantum circuit before this paper.
Other Teams Are Chasing the Same Prize
This wasn’t the only AI-for-quantum-resources story to land the same week. On November 10, The Quantum Insider also covered a separate preprint from Google DeepMind and Google Quantum AI, where AI agents paired with mathematical proofs cut the combined resource cost of five science problems, including fusion and drug-metabolism simulations, by factors ranging from about 2.5x to more than 290x. Google’s researchers hedge the same way Anthropic’s team does: these are projected costs on hardware that doesn’t exist yet, and the scientific problems still need far more resources than early fault-tolerant quantum computers are expected to have.
“The evaluator is what made this work,” says Alek Kemeny, the Anthropic physicist who mentored the project, describing why an automated correctness check mattered as much as the AI itself.
Two research groups reaching for the same lever within days of each other says something about where quantum software research is headed, even if neither result has touched real hardware yet.
Where the Evidence Stops
- No hardware run. Every check here is a classical simulation, an automated evaluator or a formal proof, not a working qubit.
- Not peer reviewed. The paper went up on arXiv on October 15, 2026. Nothing in the available write-ups says it has been reviewed since.
- A resource cut isn’t a speed cut. The combined Toffoli-times-qubit score is a proxy. Actual runtime depends on physical hardware that doesn’t yet exist at the scale these problems need.
- The hardware timeline is still distant. QuEra’s own roadmap targets its first commercial fault-tolerant system, Libra, for Amazon Braket “from 2028.” That’s a target, not a delivery date.
- Credit is genuinely blurry. Anthropic’s own write-up says Claude needed occasional prompting to look past familiar circuit designs, and that separating Claude’s contributions from its human collaborators’ “was sometimes difficult.”
- Eight examples is a small sample. Whether these gains carry over to full, real-world quantum applications isn’t shown in this paper.
Why It Matters to You
Nothing here reaches an ordinary person today. There’s no app, no cloud account, and no quantum computer quietly running a cheaper version of anything yet. The case Anthropic makes for why this matters is indirect, and it’s worth taking at face value precisely because the company doesn’t oversell it: if circuit designs like these keep getting cheaper to build, the resource wall between today’s quantum hardware and genuinely useful quantum computers could get a little lower, a little sooner. Anthropic points to new medicines, better batteries and cleaner fertilizer production as the kinds of problems cheaper quantum chemistry simulations could eventually help with, the same examples it uses to explain why resource reductions matter at all. None of those applications exist yet, and nobody in this research claims otherwise. The honest version is that a chemist, a battery engineer or a farmer won’t notice anything from this paper for years, if ever. What they might eventually notice is whether work like this keeps shaving zeros off the hardware bill standing between now and a quantum computer that solves a problem they actually have.
Watch what happens when someone tries to run one of these eight designs on an actual error-corrected chip. Toffoli counts and qubit counts are easy to improve on paper. Whether Claude’s designs hold up once real noise, real wiring and real control electronics get involved is the question nobody has answered yet, and it’s the only one that will actually settle anything.
Sources
- Anthropic: Designing More Efficient Quantum Circuits With Claude (Oct. 16, 2026)
- QuEra Computing: AI Is Cutting the Cost of Quantum Algorithms (Oct. 16, 2026)
- arXiv: Discovering State-of-the-Art Quantum Circuits With Large Language Models (submitted Oct. 15, 2026)
- The Quantum Insider: QuEra, Harvard and Anthropic Use AI to Reduce Quantum Circuit Resource Requirements (Nov. 10, 2026)
- The Quantum Insider: Google Researchers Use AI and Mathematical Proofs to Cut Quantum Computing Requirements (Nov. 10, 2026)
