Anthropic says Claude Mythos 5 improved on the best published quantum circuits
Against the smallest previous elliptic-curve design, one circuit used 12% more qubits but 17 times fewer gates; the post cites no outside expert review.
- Models & capabilities
- Ideas & essays
- Minor
Anthropic reported that Claude Mythos 5 had designed quantum circuits for eight computational primitives that used fewer qubits and gates than the best previously published designs. The work was led by Andi Gu of Anthropic’s STEM Fellows Program and mentored by Alek Kemeny, a quantum physicist at the company, with collaborators at Harvard and QuEra Computing. A preprint was posted to arXiv a day earlier.
The method was a closed loop. Claude was given plain-language specifications, proposed a circuit, and received a score from an automated evaluator that checked correctness and rated the design by qubits multiplied by gates, with no human input between rounds. The primitives covered chemistry, optimisation, quantum simulation, number theory and numerical computing. The arXiv abstract is more specific than the blog post: in seven of the eight tasks the discovered circuits significantly reduced both Toffoli-gate and qubit counts against the state of the art, and in the eighth the search found a new reversible architecture for modular inversion.
Anthropic described two results. In the first, a 53-hour run on elliptic-curve point addition produced a reversible form of the classical divsteps algorithm that avoids storing its branch history. Claude proved algebraic identities, checked them in simulation on 256-bit numbers and formalised the claim in Lean. One circuit in the resulting family used 12% more qubits than the smallest earlier design but 17 times fewer gates. In the second, a redesigned decoder for a quantum optimisation algorithm, decoded quantum interferometry, cut the decoder’s gate count by a factor of 12 relative to Google’s earlier construction.
The limits are in the post itself. Claude was conservative, repeatedly rechecking well-supported results and favouring familiar approaches, so the team required several qualitatively different approaches in parallel and occasionally prompted it to broaden its search; the “no human input” applies between rounds, not to the setup. Verification rested on the automated evaluator, Claude’s own simulations and Lean, and the post does not cite review by outside quantum researchers. The authors also note that circuit design is only one part of building a fault-tolerant machine.
The work follows the cryptography findings Anthropic reported in July.
In the commentary
What people were saying around this time — external links, from the record's commentary rail.
- 17 October 2026 · Zvi Mowshowitz · Don't Worry About the VaseLeaked Anthropic Draft AI Regulatory Bill Is Unacceptable and Troubling
- 15 October 2026 · James Ball · TransformerOpenAI’s spy shows it needs to grow up, fast
- 14 October 2026 · Steve Newman · Second ThoughtsAnthropic’s Scion: A Four Year Project in Five Months