Anthropic says its Claude Mythos 5 model has designed a quantum circuit that requires roughly 17 times fewer gates than the smallest previously known design. The company presented the result as evidence that artificial intelligence could help reduce the cost of quantum computing by improving algorithms rather than relying solely on advances in hardware.

The work was conducted under the guidance of researcher Andi Gu as part of Anthropic's STEM Fellows Program, in collaboration with researchers from Harvard University and QuEra Computing. Claude Mythos 5 took part in eight circuit design experiments spanning fields including chemistry, physics simulation, optimization, number theory, and numerical computing.

New from our STEM Fellows Program: Designing more efficient quantum circuits with Claude. Given plain-language instructions, Claude Mythos 5 designed quantum circuits that improve on the best published ones, sometimes by an order of magnitude or more. More efficient circuits…

— Anthropic (@AnthropicAI) October 16, 2026

How AI-Designed Circuits Fit Into Quantum Computing

Quantum computers perform computations on qubits through operations known as quantum gates. Unlike classical bits, which hold a state of either zero or one, qubits can occupy combined states through superposition. That property may allow quantum machines to solve certain scientific problems that are intractable for traditional computers.

Current quantum hardware, however, remains limited by qubit availability and environmental noise. Minor fluctuations introduce errors, and longer computations become increasingly difficult to carry out. Finding circuits with minimal operation and memory requirements is therefore a central task, making algorithm development an essential step toward improving quantum computing.

Anthropic supplied Claude with plain-language descriptions of eight computation tasks along with an automated system for evaluating the results. The model generated circuits, assessed its own outputs, and refined its solutions in iterative cycles without any human intervention during them. According to the company, some of the results significantly surpassed the best-known solutions by an order of magnitude.

Cutting Memory Requirements in Modular Inversion

One major result involved modular inversion, a mathematical operation used in elliptic-curve arithmetic. The operation is relevant to quantum algorithms targeting the elliptic-curve discrete logarithm problem, which has implications for evaluating the security of elliptic-curve cryptography against sufficiently capable quantum computers.

Claude developed a reversible algorithm for the arithmetic using a method known as divsteps, which simplifies numbers through consecutive conditional operations. Converting the procedure into a quantum circuit raises a memory problem: a reversible calculation must retain enough information to restore earlier steps rather than discarding unneeded temporary results.

The model discovered a technique for recovering branching choices from intermediate numerical values. It identified the relevant mathematical identities, tested the approach with numerical simulation, and proved its correctness using the Lean proof assistant, a tool that mechanically verifies mathematical proofs. As a result, the circuit does not need to store the full history of the choices made.

Fewer Gates, More Qubits

The optimized elliptic-curve circuit uses about 17 times fewer gates than the smallest previously existing design while employing 12% more qubits. The finding illustrates an important engineering trade-off: reducing the number of operations can increase memory needs, so both resources must be weighed during circuit design. Those extra qubits carry real weight on current hardware, where qubit availability is itself limited.

Anthropic also reported other improvements. One optimization cut the number of gates in a Reed–Solomon decoding circuit used in the DQI (Decoded Quantum Interferometry) quantum optimization algorithm by about 12 times. The company suggested that AI-assisted design can help identify possible reductions across a range of mathematical and scientific tasks, though every design must still be analyzed individually.

These achievements do not imply that existing quantum computers can already perform any computation required of them. Serious technical challenges remain around hardware constraints, error correction, and operational stability. The elliptic-curve finding is significant for cryptographers, but it does not prove that modern devices can break widely used cryptography.

Building on Years of Algorithmic Progress

The research builds on years of progress in reducing the estimated resources needed for demanding quantum simulations. According to Anthropic, FeMoco — an iron-sulfur cluster associated with biological nitrogen fixation that has long served as a benchmark for such resource estimates — was estimated in 2017 to require 50 trillion Toffoli gates to compute its ground-state energy. Algorithmic innovations have since brought that figure down to 340 million gates, more than 150,000 times lower than the initial estimate.

Automated optimization of circuits could help push that number lower still, although actual results will depend on hardware advancements and other factors in the future.

Source: Tron Weekly