1. Home
2. News
3. Anthropic

Anthropic says Claude designed quantum circuits that beat the best published ones on eight tasks

One design for the quantum attack on elliptic-curve keys needs 17 times fewer gates than the best one published, after a 53-hour run.

By Think Facility · October 17, 2026 · Anthropic Claude Save

On October 16, 2026, Anthropic published research saying Claude Mythos 5 designed quantum circuits that beat the best published ones on all eight tasks it was given. The researchers handed it a plain-language description of each circuit and an automated checker that scored every attempt. The model did most of the rest on its own.

The headline result is on elliptic-curve point addition, the core step in the quantum attack on elliptic-curve cryptography. After a single 53-hour run, Claude produced a design that uses 12% more qubits than the smallest previous one but needs 17 times fewer gates. A second design made the costliest step of a leading quantum optimization algorithm about 12 times cheaper.

Who

Andi Gu, with Alek Kemeny, Susanne Yelin and Mikhail Lukin

Model

Claude Mythos 5

Tasks

eight circuits from chemistry, optimization, physics simulation and number theory

Longest run

53 hours with minimal human input

Point addition

12% more qubits, 17x fewer gates

Decoder

gates cut by a factor of 12

What did Claude change?

Quantum circuits play by two rules ordinary code doesn't. They can't branch on a value, and they can't throw scratch work away, because every step has to be able to run backward. Undoing intermediate results costs gates. Keeping them costs qubits.

For point addition the expensive part is modular inversion. Ordinary computers do it with a method called divsteps, but a quantum version had to store a record of every choice the method made, and that ate qubits. Claude's idea was to read the record back out of the numbers themselves after each round. Its first rules held on small cases and broke on 256-bit numbers. So it proved new algebraic identities, derived an exact decoding rule, and checked it with large simulations and then a formal proof in Lean.

The decoder result came about the same way. Claude dropped the method behind Google's previous circuit and built on Berlekamp-Massey decoding instead. Then it wrote a simulator and ran 403 million sample decodings to see how much memory the rare cases really need. Planning for the worst case would take 665 qubits. Claude set aside 19, and about one run in 10,000 fails, which Anthropic says is within the error rate the task allows.

How does that compare with people doing it?

Anthropic measures it against a chemistry problem. In 2017, researchers estimated a quantum computer would need about 50 trillion Toffoli gates to work out the ground-state energy of FeMoco, the iron-sulfur cluster bacteria use to turn nitrogen from the air into ammonia. A run of later papers brought that to about 340 million. None of that drop was Claude's.

Claude's elliptic-curve design came out of one run lasting a little over two days. That's the comparison I'd hold on to. It's research-grade algorithm design, done in days and checked by machine at every step.

Anthropic is plain about where the model held back, too. It says Claude was "conservative by nature" and kept double- and triple-checking results that already held up, and it leaned toward well-known approaches. The team made it develop several different approaches in parallel and nudged it now and then to search more widely.

What it means for the encryption people use

Point addition matters outside the lab because of what it attacks. Anthropic describes it as the core step in quantum algorithms for the discrete logarithm problem on elliptic curves, which ordinary computers can't solve for large curves and a big enough quantum computer could. ECDSA, the signature standard built on those curves, is the same one whose NIST seeds an Anthropic model traced the same day.

Governments have been preparing for this for years. NIST finalized its first three post-quantum encryption standards on August 13, 2024, and its mathematician Dustin Moody told system administrators to start integrating them immediately, "because full integration will take time." On November 12, 2024, NIST published a draft plan for moving agencies and industry off quantum-vulnerable algorithms.

Anthropic's post doesn't say when a machine could run this design against a real key, and it gives no machine size for that. It does say the cuts reduce what useful applications need, so practical quantum computers could arrive sooner than expected. Every gate and qubit saved lowers the size of machine an attacker would have to build (and the size a chemist would need, too).

That second use is the one Anthropic leads with. It says understanding FeMoco could help design better industrial catalysts, and it lists new medicines, better batteries and cleaner ways of making fertilizer as problems faster circuit design could bring closer.

Andi Gu led the work through Anthropic's STEM Fellows Program, and the full paper is on arXiv with Harvard's Susanne Yelin and Mikhail Lukin, who is also at QuEra Computing.