Claude Mythos 5 designs quantum circuits that beat published records

Anthropic has shown that Claude Mythos 5 can independently design quantum circuits that beat the best published human solutions. In some cases the gain was an order of magnitude or more. The model received descriptions of eight circuits from chemistry, optimization, physical systems simulation and number theory, along with an automated evaluator that checks and scores each candidate. Andi Gu led the project as part of the Anthropic STEM Fellows Program. Anthropic quantum physicist Alec Kemeny served as mentor, and the co-authors were Susanne Yelin of Harvard and Mikhail Lukin (Harvard, QuEra Computing).

The bottleneck is circuits, not hardware

Today's quantum machines have hundreds or thousands of qubits. For many algorithms with practical value, that is not enough. Qubits are also sensitive to noise: heat, vibration and even cosmic rays shift their state, small errors accumulate, and the computation returns a wrong answer. Researchers therefore try to rewrite algorithms so that they need fewer qubits and less time.

A clear example is FeMoco, an iron-sulfur cluster in the enzyme nitrogenase, which bacteria use to convert atmospheric nitrogen into ammonia. Understanding how it works would help design better industrial catalysts. In 2017, calculating the ground-state energy of FeMoco was estimated to require roughly 50 trillion Toffoli gates. Today the estimate has dropped to about 340 million, a reduction by a factor of 150,000. That corresponds to several hours of runtime on a future machine with about three million physical qubits. The reduction came mainly from redesigning circuits, not from better hardware.

A closed loop without a human

The authors give three reasons for choosing Claude. First, quantum circuits are described in code, so the model can write, run and edit them like ordinary programs. Second, the class of algorithms in question is easy to verify: each input maps to exactly one output, and a conventional computer can check it. An external evaluator program therefore validates the model's proposals. Third, there is an objective metric: the number of qubits multiplied by the number of gates. It serves as a proxy for the circuit's runtime.

Claude Mythos 5 worked in a loop. It proposed a design, received feedback and a score from the evaluator, and took both into account in its next attempt. No human intervened between rounds. The tasks covered eight computational primitives, the basic building blocks of larger quantum algorithms, for chemistry, optimization, quantum simulation, number theory and numerical computation.

Key parameters

• Model: Claude Mythos 5, working in a closed loop with no human involvement between rounds
• Tasks: 8 computational primitives from chemistry, optimization, physics, number theory and numerical methods
• Input: a plain-language circuit description plus an automated evaluator
• Metric: number of qubits × number of gates
• Result: designs better than the best published ones, in places by an order of magnitude or more
• Context: the estimate for FeMoco fell from 50 trillion to about 340 million Toffoli gates

What changes for the industry

Anthropic's main conclusion is that lower resource requirements bring practical quantum computers closer than previously expected. Teams that build quantum hardware or estimate when applications become feasible now have a reason to recalculate their resource forecasts.

The approach comes with a condition. It works where an ordinary computer can verify the result and quality can be reduced to a single number. According to the authors, that condition holds for this class of algorithms. The template is portable: a natural-language task, a programmatic evaluator, one metric and an autonomous improvement loop. Details on each of the eight circuits are in the full paper on arXiv.