Quantum computers will only become useful when two curves meet: the hardware gets bigger and more reliable, and the algorithms that run on it get cheaper. A new collaboration between QuEra Computing, Harvard University and Anthropic targets the second curve, and the numbers are eye-catching. According to QuEra’s November 10, 2026 announcement, AI-designed circuits cut a combined gate-and-qubit resource metric by as much as 39 times.
What the researchers actually did
The work is described in a preprint titled “Discovering state-of-the-art quantum circuits with large language models” (arXiv:2610.22367). It looks at eight computational building blocks that appear inside larger quantum algorithms. In seven of them, circuits designed with AI used fewer gates and fewer qubits than the baseline designs. In the eighth, the search turned up a new approach that gives engineers additional ways to trade qubits against gates.
The project was led by Andi Gu as part of Anthropic’s STEM Fellows Program, with Alek Kemeny of Anthropic, Susanne Yelin of Harvard, and Mikhail Lukin, who works at both Harvard and QuEra. The setup paired Anthropic’s Claude with an independent automated evaluator. The evaluator checked each proposed circuit for correctness and scored it on two numbers: its Toffoli-gate count and its qubit count. Claude used that feedback to keep exploring and improving.
That pairing is the interesting part. Toffoli gates can account for a large share of the overhead in fault-tolerant computing, while logical qubits take up scarce hardware capacity. Because every candidate was independently verified, the model could try unconventional ideas and the team could trust the ones that survived.
The headline results
QuEra highlights three examples from the paper:
- Optimization: the Reed-Solomon decoder, the most expensive step in decoded quantum interferometry, dropped from 5.1 million Toffoli gates to about 440,000, a 12x reduction, while also using fewer qubits.
- Chemistry: for the Coulomb potential term used when simulating molecules, Toffoli gates fell from 6,355 to 1,140 and qubits from 252 to 142, a 9.9x cut in the combined metric.
- Numerical computation: a fixed-point Gaussian function went from about 24,000 Toffoli gates and 362 qubits to 1,583 gates and 139 qubits, the 39x reduction that gives the announcement its headline.
The authors are careful about what those figures mean. The combined metric multiplies Toffoli count by peak logical qubits, so a 39x improvement is not a 39x speedup. The results are resource estimates for individual building blocks, not runs on quantum hardware, and their effect on a complete application depends on how the pieces are combined.
A new subroutine found by AI
Beyond tuning existing designs, Claude helped discover a new reversible approach to modular inversion, an arithmetic operation used in elliptic-curve calculations. Reversibility makes many efficient classical tricks expensive on a quantum machine, because intermediate information normally has to be kept around. The new method reconstructs much of that information instead of storing it, which opens new options for balancing gates against qubits.
In a full point-addition benchmark, one design needed roughly 17 times fewer gates than the previous design that used the fewest qubits, at the cost of about 12 percent more qubits. QuEra says a 53-hour search, guided by researcher-defined requirements and the automated evaluator, produced this approach.
From algorithms to everyday operation
The release also ties the circuit work to an earlier QuEra and Anthropic pilot on laser control. Stable lasers are essential for running QuEra’s machines. 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 AI. QuEra sums up three roles for AI in quantum computing: improving algorithms, supporting reliable operation, and developing control software.
What it means
Algorithm progress has historically moved the goalposts for hardware. QuEra points out that successive advances cut the estimated Toffoli-gate count for one widely studied FeMoco simulation by roughly 150,000 times over eight years. If AI-assisted search can speed up that kind of work, near-term fault-tolerant machines may be able to tackle larger problems than their raw size suggests.
For readers outside quantum, the broader lesson is the method: open-ended AI search combined with a rigorous, independent checker. Kemeny of Anthropic called that pairing “a promising model for AI-assisted discovery well beyond quantum computing.” QuEra says its Libra fault-tolerant system is targeted for availability on Amazon Braket in 2028, so the practical payoff remains some distance away, and these are preprint-stage estimates rather than hardware demonstrations.
Conclusion
The takeaway is not that AI has solved quantum computing. It is that a carefully verified AI loop can find smaller circuits than human-designed baselines, at least on paper, and that could shorten the road to useful quantum applications.