Anthropic says its Claude Mythos 5 AI can design quantum circuits that beat existing benchmarks on seven out of eight standard quantum tasks. According to the company, these designs use fewer qubits and gates than anything published before—sometimes by more than a factor of ten. If that holds up, it could really cut down the quantum resources needed for all sorts of problems in chemistry, optimization, materials science, and cryptography.
Here’s how they did it: Anthropic gave Claude plain-language briefs for tasks in areas like chemistry, optimization, physics, and number theory. For each one, an automated checker tested if the circuit was correct, and then scored it by how many qubits and gates it needed.
Claude worked in a closed loop: it coded a solution in Python, the checker gave feedback, and Claude made tweaks—all without humans lifting a finger in-between. In the end, Anthropic says the AI found completely new designs, showing it could slash both qubit and gate requirements compared to the best-known algorithms.
A quick refresher: Quantum circuits are basically just sequences of operations (gates) applied to qubits. Since today’s quantum computers are small and pretty noisy, shaving off even a few gates or qubits can make a big difference—costs go down and running the algorithm becomes more realistic.
One result stands out: elliptic-curve point addition, which matters for quantum attacks on the discrete logarithm problem. Claude managed to create the first reversible version of the “divsteps” modular-inversion technique that skips storing branch history—a clever fix for a notorious memory bottleneck. To be clear, their design does use 12% more qubits than the previous best, but it needs 17 times fewer gates. That’s not an outright win on every front, but it gives researchers a new way to balance memory versus speed.
Claude also rebuilt part of the decoding process in quantum interferometry, swapping out a previous algorithm with Berlekamp-Massey decoding. That cut the decoder’s gate count by a factor of twelve compared to a design proposed by Google Quantum AI.
The project was led by Andi Gu at Anthropic under their STEM Fellows program, with guidance from quantum physicist Alek Kemeny and input from Harvard researchers Susanne Yelin and Mikhail Lukin (who’s also at QuEra Computing). For evaluating Claude’s solutions, they kept it simple: just minimize the product of qubit count and gate count—a basic way to track real circuit cost. Anthropic insists Claude’s advances came from fresh structural and mathematical insights, not by tweaking older designs.
During the elliptic-curve work, Claude spent 53 hours experimenting, mostly without human help. It tried various strategies, tested its own rules, and even formalized some reasoning in the Lean proof assistant.
Why does this matter? If these results hold up, AI like Claude could seriously speed up one of the slowest, most painstaking parts of quantum algorithm development. For comparison, getting the gate cost for the FeMoco molecule (which matters for nitrogen fixation and fertilizers) down from 50 trillion Toffoli gates in 2017 to about 340 million took several research groups years. Anthropic thinks faster circuit design could accelerate everything from catalyst discovery to better batteries and greener agriculture. They also suggest that similar AI techniques could help with error correction and hardware design too.
But here’s the catch: this is all based on a company research report and a preprint—not a peer-reviewed paper yet. The circuits have only been tested with simulators and formal checks, not on real, large-scale fault-tolerant quantum computers.
Anthropic admits there were some rough edges. Claude tended to play it safe, often re-confirming answers it already understood, and leaning toward familiar strategies. The team fixed this by getting the AI to try more approaches in parallel and nudging it to explore new territory now and then.
So yes, it’s a big deal in AI-driven quantum circuit design, at least as far as new research goes. But the real impact? That depends on whether other scientists can replicate these results, validate them through peer review, and, eventually, see them work on actual quantum hardware.
