Claude Mythos 5 beat prior state-of-the-art circuit designs on seven of eight tasks with minimal human help
Quantum computers have a resource problem. The algorithms people want to run often demand more operations and more qubits than today’s machines can comfortably handle.
Anthropic says its Claude AI just made a dent in that problem. On October 16, 2026, the company announced that its Claude Mythos 5 model autonomously designed quantum circuits that beat prior state-of-the-art designs on seven of eight benchmark tasks, in one case using up to 17 times fewer gates.
A quantum circuit is the recipe a quantum computer follows. It is a sequence of logic operations applied to qubits, the quantum version of the bits inside a regular computer. Shorter recipes with fewer ingredients are easier to run on real hardware.
Anthropic pointed Claude at eight computational tasks drawn from different scientific fields. The list spanned chemistry, optimization, physics simulation and number theory.
The setup was a closed-loop evolutionary framework. Claude proposes a circuit, a judge grades it, and Claude uses the feedback to try again.
Claude received each task as a plain-language specification. It then wrote and refined circuit designs on its own. After the initial prompts, the model operated with minimal human intervention.
The judge was an automated evaluator with two jobs. First, it checked that each circuit actually produced correct results. Second, it scored designs on how many Toffoli gates and qubits they used.
A Toffoli gate is a three-qubit logic operation, and its count served as one of the main cost metrics here. Fewer Toffoli gates and fewer qubits meant a better score.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
In seven of the eight tasks, Claude’s circuits came in under the previous state-of-the-art designs. Those designs reduced both Toffoli gate counts and qubit requirements.
The standout was a modular inversion architecture. That design achieved up to 17 times fewer gates than earlier approaches.
There was a catch, though a modest one. The modular inversion circuit required slightly more qubits to pull off its gate savings.
The project was led by researcher Andi Gu through Anthropic’s STEM Fellows Program. Gu worked alongside collaborators from the Harvard Quantum Initiative and QuEra Computing.
The approach leans on what large language models already do well: writing code and reasoning through problems step by step. The automated evaluator adds the rigor, so a confident-sounding but wrong circuit never makes the cut.
Current quantum hardware is noisy and limited. Machines can only run so many operations before errors pile up, and qubits remain a scarce resource.
For the FeMoco chemistry problem, a molecular simulation target, earlier human-led advances cut estimated resource requirements from roughly 50 trillion Toffoli gates to about 340 million. Claude’s work suggests AI could become another tool in that same optimization effort.
For the quantum computing industry, the most immediate takeaway is about process, not any single circuit. If an AI model can iterate on circuit designs with little human input and beat specialist benchmarks, research teams gain a way to explore far more design options than people could test by hand.
There are clear caveats. These were eight selected tasks, and one of them did not produce an improvement over existing designs. The modular inversion gains also came with a qubit tradeoff, which matters on hardware where qubits are scarce.
For Anthropic, the project strengthens its pitch that frontier AI models can contribute to hard scientific work. The STEM Fellows Program framing suggests the company wants Claude seen as a research collaborator, not only a chatbot or coding assistant.