Claude designs quantum circuits with 17 times fewer gates

What this is about

Anthropic and Harvard University researchers reported on October 16, 2026 that Claude Mythos 5 improved eight quantum circuits. The model received a description of the desired computation and then worked with an automated evaluator. The evaluator checked whether a design produced the correct result and scored its qubit and gate counts.

The strongest single figure comes from a circuit for elliptic-curve point addition. After a 53-hour run, Claude's design required 12 percent more qubits than the smallest earlier approach, but 17 times fewer gates. This is a computational resource comparison, not a demonstration on a quantum computer.

What Claude's quantum-circuit search actually does

Claude wrote circuit designs as Python code. An external program checked each proposal, assigned a score and returned the result to the model. Claude then changed its approach and repeated the process. No human intervened between those rounds; researchers supplied the initial task and occasionally asked for broader solution paths.

The eight tasks came from chemistry, optimization, physics simulation, number theory and numerical computing. In a second highlighted case, Claude replaced one decoder with a different mathematical method and reduced expensive operations by roughly a factor of 12. Parts of the mathematical argument were formalized in Lean, while other results were checked through simulations and sampling.

The approach works especially well because these circuits can be checked efficiently on classical computers. The model can try many variants while the evaluator rejects incorrect results immediately.

Why it matters

Today's quantum computers have too few stable qubits for many useful algorithms. Every additional gate also adds runtime and more opportunities for errors. Better circuits can therefore matter as much as better hardware.

Anthropic gives FeMoco molecular simulation as a historical comparison: algorithmic advances reduced estimated costs from about 50 trillion Toffoli gates in 2017 to roughly 340 million. That history shows how much smarter circuits can change hardware requirements. Claude's results are smaller and more narrowly scoped, but follow the same logic.

The method is also instructive for AI research. Long autonomous runs become more credible when an independent evaluator scores every intermediate solution. This differs from a chatbot answer that merely sounds plausible: there is a concrete mathematical specification and a measurable resource cost.

In plain language

Imagine packing a suitcase for a trip. After every attempt, someone checks whether all required items are inside, how heavy the suitcase is and how many actions packing took. Claude can pack again and again and receives only those measurements. After many attempts, it finds an arrangement that takes slightly more space but needs far fewer actions. That is the trade-off between qubits and gates studied by the researchers.

A practical example

A research team wants a circuit that will be repeated thousands of times inside a future quantum application. Its existing design needs 200 qubits and 1.7 million gates. An AI-assisted search proposes a version using 224 qubits and 100,000 gates. Memory demand rises by 12 percent while the operation count falls to one seventeenth.

Before adopting the new design, the team must do more than compare those two numbers. It tests large inputs, investigates rare failures, accounts for the target hardware's connectivity and calculates error-correction overhead. Only then can it determine whether the advantage survives on a real machine.

Scope and limits

• The work is a public preprint and has not completed standard peer review. Independent replication is still missing.
• The circuits were checked computationally but not run on a real fault-tolerant quantum machine. Gate and qubit counts are only proxies for actual runtime.
• One decoder accepts a reported failure probability of about one in 10,000. That may be unacceptable for some applications.
• The method needs tasks with a reliable automated evaluator. It does not transfer easily to open research questions without a clear scoring function.
• Fewer gates do not mean current encryption has been broken. Large fault-tolerant quantum computers do not yet exist.

SEO & GEO keywords

Claude Mythos 5, Anthropic, quantum circuits, quantum computers, qubits, quantum gates, Harvard University, elliptic curves, Lean, formal verification, AI for science