AI Is Now Designing Quantum Circuits That Need Far Fewer Qubits and Gates

Useful quantum computing depends on two things arriving at the same time: hardware powerful enough to run the machines, and algorithms lean enough to run on them. A new collaboration between Harvard University, Anthropic and QuEra Computing has moved the second of those curves, showing that artificial intelligence can design quantum circuits that need far fewer resources than the best human-designed ones.

The research, posted as a preprint to arXiv on October 15, 2026, and highlighted by QuEra on November 10, used Anthropic's Claude model to design circuits for eight computational building blocks — the small, repeated operations that sit inside larger quantum algorithms, spanning numerical computation, chemistry, optimization, quantum simulation and number theory.

In seven of the eight tasks, the AI-designed circuits reduced both the number of gates and the number of qubits needed compared with the best published designs. In the eighth, the search found a new approach that offers fresh choices for balancing one against the other.

How the AI did it

The method paired Claude with an independent automated evaluator that checked every proposed circuit for correctness and scored how many resources it used. Claude used that feedback to explore and improve its designs, in an evolutionary search. Because every design was independently checked, the model could pursue unconventional ideas and the researchers could trust the ones that held up.

The work was led by Andi Gu as part of Anthropic's STEM Fellows Program, with Alek Kemeny of Anthropic and Susanne Yelin and Mikhail Lukin of Harvard. Lukin is also a co-founder of QuEra.

The headline numbers — and what they actually mean

The largest improvement came in a numerical computation task, a fixed-point Gaussian function, where the AI-designed circuit achieved a 39-fold reduction in the combined resource metric, with Toffoli gates falling from about 24,000 to 1,583 and qubits from 362 to 139.

In optimization, the Reed-Solomon decoder — the most expensive step in a promising quantum optimization algorithm called decoded quantum interferometry — needed 12 times fewer Toffoli gates, dropping from 5.1 million to about 440,000, while also using fewer qubits. In chemistry, the Coulomb potential calculation, which describes how charged particles interact when simulating molecules, improved 9.9-fold on the combined metric, with Toffoli gates down from 6,355 to 1,140 and qubits from 252 to 142.

One number deserves care. The combined metric multiplies the Toffoli-gate count by the peak number of logical qubits, so a 39-fold improvement in that metric is not a 39-fold speedup. Toffoli gates are the expensive, hard-to-perform operations in fault-tolerant quantum computing, and logical qubits occupy valuable hardware capacity, so reducing both matters — but the effect on a complete application will depend on how those components are used.

A genuinely new idea, not just a tweak

Beyond improving existing designs, Claude helped discover a new reversible approach to modular inversion, an arithmetic operation central to elliptic-curve calculations. The approach reconstructs much of the information normally stored during intermediate steps, creating new options for balancing gates against qubits. In the complete point-addition benchmark, one design needed about 17 times fewer gates than the prior design that used the fewest qubits, while using about 12 percent more qubits.

That matters because reversibility makes many efficient classical methods costly to implement on quantum computers. As Mikhail Lukin put it, what is exciting is that AI, guided by a rigorous evaluator, found new circuit structures that work more efficiently within that constraint.

The search that produced the new elliptic-curve approach ran for 53 hours, guided by researcher-defined requirements and the automated evaluator.

What this does not show

Two caveats are important. First, the paper is a preprint and has not been peer-reviewed. Second, these are resource estimates for individual building blocks, checked by classical simulation — nothing was run on quantum hardware, and no machine that could run these circuits exists yet. The work does not give a new estimate for breaking encryption, and it is not a demonstration of an attack on deployed cryptography.

Still, the direction of travel is significant. QuEra points out that successive advances reduced the estimated Toffoli-gate count for one widely studied molecular simulation by roughly 150,000 times over eight years. This study suggests AI could accelerate parts of that research process.

AI is already helping run the machines

The circuit-design work builds on an earlier QuEra pilot with Anthropic on laser control. Laser stability is a key requirement for reliable operation of QuEra's neutral-atom quantum computers. In that pilot, Claude developed and validated control software that recovered the correct laser lock in 695 of 700 trials, in seconds rather than the 5 to 10 minutes a human expert needs. The resulting controller is conventional software that runs without the AI.

Together, the two projects show AI playing three roles in quantum computing: improving algorithms, supporting reliable operation, and developing control software.

QuEra is building fault-tolerant quantum computers for real work, with its Libra system targeted for availability on Amazon Braket in 2028. Libra is designed as a megaquop-class machine with 256 error-corrected logical qubits, announced in June 2026 through an expanded collaboration between QuEra and Amazon Web Services.

Sources

• PR Newswire — Harvard, Anthropic, and QuEra Breakthrough Research Uniquely Shows AI Can Reduce the Resources Quantum Applications Need
• arXiv:2610.22367 — Discovering state-of-the-art quantum circuits with large language models (Gu, Kemeny, Yelin, Lukin)
• AWS Quantum Computing Blog — AWS Deepens Strategic Collaboration with QuEra to Bring Fault-Tolerant Quantum Computing to Amazon Braket