Anthropic's AI optimizes quantum circuits

Anthropic says Claude Mythos 5 designed quantum circuits more efficiently than the best published designs to date, reducing the number of operations by a factor of 17 in some cases and that of another circuit by 12. The work focused on eight tasks in chemistry, optimization, physics simulation, number theory and numerical computation.

The idea matters because today's quantum computers still have a basic problem: they contain few qubits, the units of quantum information, and are highly sensitive to noise. Each additional operation increases the risk of error and the time needed to complete a calculation.

Claude designed, tested and improved the circuits

Anthropic's team did not ask Claude to write a theoretical answer. It gave the model plain-language instructions describing what each circuit had to calculate, along with an evaluator program that could check whether it worked and assign it a score.

The model worked in a closed loop:

• It proposed a design.
• The evaluator checked whether it was correct.
• It measured how many qubits and quantum gates the design used.
• Claude used that result to create an improved version.

There was no human intervention between most rounds. According to Anthropic, the model found designs that used fewer qubits and fewer gates than previously published algorithms. In some cases, it did not simply adjust existing circuits, but changed the way the problem was solved.

A quantum gate is a basic operation applied to one or more qubits. The more gates a circuit needs, the longer it takes and the more opportunities there are for noise to alter the result. That is why reducing them can be as important as building a processor with more qubits.

One example: calculating with less memory and less time

One of the experiments tackled modular inversion, a costly mathematical operation in calculations involving elliptic curves. This type of calculation is relevant to certain quantum algorithms that could, in theory, solve problems that are impractical for classical computers.

Claude worked for 53 hours with an initial description and some guidance to explore alternatives. The model found a way to avoid having the circuit store all the intermediate information it needed to run backward, a requirement of quantum computing.

On a conventional computer, a program can delete temporary results when it no longer needs them. In a quantum circuit, that information cannot simply be erased: it must be reconstructed by running operations in reverse. Storing it requires qubits; recalculating it requires time.

Claude discovered that some of this information could be recovered directly from the current state of the calculation. The resulting design uses 12% more qubits than the previous smallest circuit, but needs 17 times fewer gates. This expands the possible trade-offs between memory and speed instead of offering a single optimal solution.

A quantum decoder that is 12 times cheaper

The second result focused on an optimization algorithm called DQI. Its task is to increase the probability that a quantum computer will find good solutions to problems with many possible options, such as scheduling, logistics or materials design.

The most expensive step in the algorithm was a decoder based on the extended Euclidean algorithm. Claude replaced it with another method, known as Berlekamp-Massey decoding, which runs a more regular and predictable sequence. That structure fits better with the constraints of a quantum circuit.

The model also reduced the memory reserved for exceptional cases. Instead of preparing space for the worst-case scenario, it simulated 403 million decodings and found that almost all of them needed to store at most one value. The circuit reserves 19 qubits for that case and fails in approximately one out of every 10,000 executions, a rate the team considers acceptable for this task.

The result was a 12-fold reduction in the number of gates compared with Google's previous design. Anthropic says the improvement does not come from a single adjustment, but from changing the mathematical method, the circuit architecture and the amount of memory reserved.

What this means for you

This does not immediately turn quantum computers into machines available to businesses or consumers. Current hardware is still limited to hundreds or thousands of qubits, far fewer than several practical applications require, and it also needs complex systems to detect and correct errors.

What does change is the speed at which new designs can be explored. Anthropic notes that the resource estimate for simulating the FeMoco molecule, which is important for studying catalysts and fertilizer production, fell from around 50 trillion Toffoli gates in 2017 to around 340 million after years of progress by different groups. According to the company, some of Claude's new designs emerged in a single 53-hour run.

The claim still needs to be evaluated by the scientific community and on real hardware. The circuits were checked with automated evaluators and simulations, and one of the results was also formally verified in Lean. But a theoretical improvement does not by itself guarantee a practical advantage on a real quantum machine.

The next step will be to apply this approach to more parts of the system, from algorithms to error correction and hardware control. If AI can optimize them together, it could reduce one of the less visible bottlenecks in quantum computing: not just building better machines, but finding instructions those machines can execute without wasting their scarce resources.