Claude Mythos 5 designs record-efficient quantum circuits

Claude Mythos 5's headline elliptic-curve result rests on a specific mathematical trick. In reversible "divsteps" modular inversion, a quantum circuit normally has to store the sign and branch decision made at every arithmetic step, which eats up qubits. Claude found a mathematical invariant that lets those branch decisions be recovered after the fact from the updated arithmetic registers instead of being stored separately, removing the need for dedicated bookkeeping qubits. The resulting design uses about 12% more qubits than the best previous construction but cuts the gate count to roughly one-seventeenth ^[2]. Claude went further than a typical circuit paper would and formally proved the correctness of this arithmetic invariant in the Lean theorem prover, rather than relying only on spot-checks ^[2].

A second result followed a different path. For a Reed-Solomon decoder circuit used in a scheme called Decoded Quantum Interferometry, Claude swapped the standard decoding approach for Berlekamp-Massey decoding and built a custom simulator to size the circuit's memory precisely instead of over-provisioning qubits defensively. That swap cut the Toffoli gate count roughly 12x against Google Quantum AI's prior construction, from about 5.1 million gates down to about 440,000, while also trimming peak qubit use from 1,938 to 1,837 ^[3]. Anthropic separately reports that the storage optimization alone cut qubits needed for stored records from 840 down to 19 ^[1].

The framing of "minimal human input" obscures a more interesting detail buried in Anthropic's own writeup: Claude was the conservative party in this collaboration, not the reckless one. The company describes it as conservative by nature, noting it repeatedly second-guessed apparent improvements and continued to double- and even triple-check results that were already well supported ^[1]. Human researchers had to actively counteract this by prompting it to broaden its search beyond established approaches ^[2], rather than settle for safe, incremental tweaks to known designs.

The elliptic-curve modular-inversion design took 53 hours end to end, with human involvement limited to initial instructions plus occasional nudges to explore further ^[2]. Scoring itself was handled by an evaluator program kept separate from the model proposing designs, grading candidates by Toffoli-gate count multiplied by peak qubit use, so a plausible-looking but incorrect circuit would score zero rather than winning by seeming convincing ^[1].

The "minimal human input" framing itself became a point of friction once the announcement reached a wider audience: on Reddit, several commenters pushed back on the idea that the process was as hands-off as described, arguing it understates how much engineers actually steered the work, while the thread's original poster replied that Anthropic's writeup described a series of recursive improvement steps, without fully resolving the disagreement.

Headline multipliers like 17x or 12x only mean as much as the baseline they are measured against, and that baseline varies across the eight benchmarks. For at least one of the improved tasks, a Hermitian-code decoder, no previously published reversible quantum circuit existed in the literature at all; Anthropic's own comparison point was a straightforward reversible construction it built itself rather than an established prior design ^[1]. That is a meaningfully weaker form of evidence than beating a peer-reviewed baseline, and it is the specific gap an outside reader should weigh before taking all eight wins at face value.

That same instinct showed up organically in public reaction: in a widely upvoted Reddit discussion of the announcement, the top comment questioned whether the results had actually been checked against prior published literature or only against Anthropic's own benchmark, framing quantum circuits as a result that should be objectively checkable against outside work, not just self-reported gains. Separately, none of the results have been run on physical quantum hardware. The decoder's roughly 1-in-10,000 failure rate came from large-scale numerical simulation (about 403 million sample runs) ^[2], and the broader preprint, posted to arXiv around October 15, 2026, has not been peer-reviewed ^[3]. The gate-count savings are real as computational resource estimates, but they remain unproven on an actual quantum computer.

Quantum algorithms are only as useful as the circuits that implement them, and shrinking those circuits has historically taken years of human effort. The chemistry benchmark domain illustrates the scale involved: estimates for simulating the FeMoco (nitrogenase) molecule fell from about 50 trillion Toffoli gates in 2017 to about 340 million gates by 2026, a roughly 150,000-fold reduction achieved through nearly a decade of incremental algorithmic advances by the research community ^[2]. Claude's comparable elliptic-curve redesign took 53 hours of mostly autonomous work.

The decoder work carries a concrete timing stake: the Optimal Polynomial Intersection problem that the Reed-Solomon decoder addresses is estimated to take about 3 months on the Frontier supercomputer classically, versus roughly 1 hour on a future quantum computer ^[1]. That speedup is only realizable if the circuit implementing it is efficient enough to fit within the qubit budgets near-term hardware actually has. The project's collaborator list reflects that stake directly: Mikhail Lukin, one of the Harvard researchers on the team, is also co-founder and Chief Scientist of QuEra Computing ^[2], tying the circuit-efficiency work to a company with a commercial interest in practical quantum hardware.