Anthropic’s Claude Mythos 5 designed quantum circuits that outperform the best published algorithms across eight benchmark tasks. The AI model reduced gate counts by up to 17 times with minimal human input, according to research the company published on October 16.
The tasks covered chemistry, optimization, physics simulation, number theory, and numerical computing. In each case, Claude received a plain-language description and worked in an automated loop. It proposed circuit designs in Python, received feedback from an evaluator, and revised its approach without human intervention between rounds.
The largest improvement came in elliptic curve point addition, a key step in quantum cryptography algorithms. Claude worked for 53 hours and produced the first reversible implementation that stores no branch history. As a result, the design uses 17 times fewer gates than the previous best, with only 12% more qubits.
Circuit efficiency matters because quantum computers require enormous resources. For example, simulating FeMoco, a molecule key to biological nitrogen fixation, once required an estimated 50 trillion gates. Human researchers reduced that to about 340 million over eight years. Claude’s approach could accelerate similar reductions across other problems.
In addition, Claude cut gate counts by 12 times for a quantum optimization decoder called decoded quantum interferometry. Google Quantum AI previously built the leading version of this circuit. Claude achieved the reduction by switching the decoding method and reconstructing temporary values from the circuit’s current state instead of storing them.
As Anthropic puts it:
Rather than designing each component on its own, AI could optimize all of them jointly, from the algorithm down to the hardware, to build an entire error-corrected quantum computer tailored to a specific problem. Such designs could also tell hardware builders which capabilities matter most for a given application. That could put problems such as designing catalysts and understanding new materials within reach of the first fault-tolerant machines.
Claude also proved new algebraic identities during the elliptic curve work and formalized its claims in Lean, a mathematical proof language. Furthermore, the model ran 403 million simulated decodings to validate its optimization decoder design.
Andi Gu led the project through Anthropic’s STEM Fellows Program. Alek Kemeny, a quantum physicist at Anthropic, mentored the work. The team also collaborated with Susanne Yelin and Mikhail Lukin from Harvard University and QuEra Computing.
Anthropic says AI-driven circuit design could help make useful quantum computing arrive sooner. However, the company notes that circuit design addresses only one part of fault-tolerant quantum computing. Error-correcting codes, decoders, and hardware control also need improvement. Target applications include catalyst design, new medicines, better batteries, and cleaner fertilizer production.
