Claude redesigns quantum circuits and beats top benchmarks by 12x

Building a useful quantum computer isn't just a hardware problem. It is equally a resource puzzle.

Today's machines pack only hundreds to thousands of noisy qubits, far too few to tackle ambitious real-world simulations like fertilizer catalysis. In 2017, simulating the FeMoco molecule was estimated to take 50 trillion Toffoli gates. Years of clever algorithmic engineering dragged that figure down by 150,000 times, proving that better circuit design often beats waiting for better hardware.

Anthropic just showed that AI can take over that drawing board. Led by Andi Gu through the Anthropic STEM Fellows Program, researchers prompted Claude Mythos 5 with plain-language specifications for eight quantum circuit primitives across chemistry, optimization, physics simulation, and number theory. Working in an automated feedback loop, the model beat the best published benchmarks on every single one, slashing costs by up to an order of magnitude or more.

Closing the loop without human hands

Designing quantum circuits is notoriously painful compared to classical coding. Qubits cannot execute straightforward if-else branches, and because every operation must remain reversible, you cannot simply delete scratch work. You have to actively "uncompute" intermediate values using extra gates, forcing brutal trade-offs between qubit memory and gate runtime.

Anthropic exploited a structural opening: these specific quantum primitives can be evaluated deterministically on classical machines. The team built an autonomous closed loop where Claude proposed circuit code, an evaluator verified correctness, and a scoring function calculated the product of qubits and gates. Claude digested the score and iterated with zero human guidance between rounds.

Over a 53-hour continuous run, the model devised a novel modular inversion routine for elliptic-curve point addition, fundamentally re-architecting how branch memory is handled. In a leading quantum optimization algorithm, Claude made the single most expensive computational step roughly 12 times cheaper. On molecular simulations, it spotted natural geometric symmetries on its own, reusing repeated calculations and gutting costly gates that only existed to handle edge cases.

Why it matters

We think this is one of the cleanest demonstrations yet of AI pushing the frontier of exact science.

Most current AI research assistants are glorified literature search engines or basic code copilots. Anthropic succeeded here because quantum circuit optimization offers an ideal sandbox: the problem space is mathematically vast and counterintuitive to humans, yet the ground truth is strictly verifiable via classical simulation. By pairing LLM reasoning with automated evaluators, the team eliminated hallucination from the final deliverable.

For the quantum industry, this shifts the battle lines. Hardware advances remain slow, expensive, and constrained by physical noise. Trimming gate counts by 12x via software is equivalent to years of hardware roadmaps arriving ahead of schedule. While these benchmarks focus on isolated primitives rather than full-scale production workflows, it is now clear that frontier models can spot novel mathematical structures that human theorists missed.