Claude Mythos 5 Designs Leaner Quantum Circuits

Working with Harvard and QuEra, Anthropic reports major gate reductions via autonomous code loops and formal proofs.

Anthropic published a technical report on Friday (October 16, 2026) claiming its Claude Mythos 5 model outperformed published human designs when creating quantum circuits. According to the company and a preprint posted to arXiv, the system beat established benchmarks across seven of eight core routines used in physical simulation, optimization, and cryptography. In specific subroutines, efficiency gains in the product of qubits and logic gates reached up to 17 times.

How quantum circuits work and where hardware pinches

Standard computers process data in bits that take values of 0 or 1. Quantum machines rely on qubits, units that exploit quantum mechanics to evaluate multiple calculation paths at once. Every instruction applied to these qubits is called a quantum gate. Complex calculations depend heavily on Toffoli gates, three-qubit logic operations that carry heavy execution costs.

The central problem for quantum computing remains hardware stability. Today’s physical qubits are fragile and lose information when exposed to tiny shifts in temperature or vibration. Moreover, quantum circuits require reversible computing: information cannot simply be overwritten or cleared mid-calculation, requiring extra steps to uncompute intermediate data. Until processors with millions of fault-tolerant physical qubits exist, squeezing theoretical circuits down to their absolute minimum is the only way to run practical algorithms.

53 hours of code and autonomous verification

The research was led by Anthropic fellow Andi Gu, advised by physicist Alek Kemeny, in collaboration with Susanne Yelin from Harvard University and Mikhail Lukin from Harvard and QuEra Computing. The model ran inside a closed loop. Researchers provided plain-text specifications for eight mathematical primitives and linked Claude directly to an automated evaluation harness.

Working from those instructions, Mythos 5 generated Python scripts, compiled candidate designs against the test harness, read error outputs, and patched its own code without human intervention. When tasked with modular inversion, a bottleneck when adding points on elliptic curves for cryptography, the model ran uninterrupted for 53 hours. Along the way, it developed a new reversible computing structure and verified its algebraic proofs using the Lean formal assistant.

Beating Google designs and the missing hardware test

On another benchmark focused on the Discrete Quantum Inference (DQI) algorithm, Claude replaced the standard extended Euclidean method with the Berlekamp-Massey algorithm. To size helper qubits, the system ran 403 million Monte Carlo simulations inside a classical simulator. Anthropic states that the final circuit cut gate counts by 12 times compared to a design previously published by Google Quantum AI, a commercial rival.

The benchmark results come with practical boundaries. None of these circuits have run on physical fault-tolerant quantum chips, which remain experimental in research labs. Every reported gain was measured through theoretical counts and classical simulation. External researchers are now examining whether these theoretical gains survive when mapped onto the physical geometric limits of real hardware. Furthermore, Mythos 5 is not open to general developers; Anthropic restricts access to partners in its cybersecurity and biological science programs.