Quantum computers are short of two things: qubits and time. Today's machines have hundreds to thousands of qubits, and those qubits are fragile, so much of the progress towards useful quantum computing has come not from hardware but from redesigning the circuits algorithms run on. Anthropic reports that Claude Mythos 5 can now do some of that redesign itself, and in its tests it beat the best published designs.
The setup was a closed loop. Claude received a plain-language brief describing what a circuit must compute, wrote a design as code, and got back a correctness check and a score for the qubits and gates it used, with no human input between rounds. Across eight building blocks used in chemistry, optimisation, quantum simulation, number theory and numerical computing, Anthropic says it found designs that used both fewer qubits and fewer gates than the best published versions. The gains came from new structure, not tuning; in simulating an iron-sulphide molecule, for example, Claude noticed that the molecule's symmetry let some results be computed once and reused.
Two results stand out. For elliptic-curve point addition, a standard benchmark and the core step in quantum attacks on elliptic-curve cryptography, Claude worked for 53 hours from a one-paragraph brief and adapted divsteps, a classical method for modular inversion that had never yielded a practical quantum circuit because it needs to keep a record of every branch. Claude found a way to infer that record from the numbers themselves. One circuit in the resulting family uses 12% more qubits but 17 times fewer gates than the smallest earlier design, and the formula was checked by large simulations and then formalised in Lean. For decoded quantum interferometry, an optimisation algorithm, it cut the gates in the decoder 12-fold against Google's previous construction.
Context helps. Algorithmic advances have already cut the estimated gate count for simulating FeMoco, the core of the enzyme bacteria use to fix nitrogen, from about 50 trillion in 2017 to about 340 million. Automated circuit search could speed up that kind of work, and the point-addition result is a reminder that the same progress that helps chemistry also lowers the cost of breaking today's cryptography on future machines.
The caution is that this is Anthropic's own write-up. Some designs are backed by Lean proofs and others by simulation, and none has been run on quantum hardware, which does not yet exist at the needed scale. Anthropic notes that Claude tended to prefer familiar approaches until pushed, and that circuit design is only one piece of a working quantum computer, alongside error correction and control hardware. Independent checks by quantum researchers will matter most.