Claude Just Redesigned Quantum Circuits Up to 17 Times Leaner — AI for Science Found a Problem It Can Actually Check
The model did not write about the science. It did a piece of it
Most AI-for-science headlines are a model summarising papers faster. Anthropic’s result, published Friday, October 16, is a different kind of claim: its Claude Mythos 5 model was given plain-language specifications for eight quantum circuits, left in a closed loop with an automated evaluator, and came back with designs that beat the best previously published versions on most of the tasks — in one case using 17 times fewer gates.
The work came through Anthropic’s STEM Fellows Program. It was led by researcher Andi Gu, mentored by Anthropic quantum physicist Alek Kemeny, and produced with Harvard’s Susanne Yelin and Mikhail Lukin of Harvard and QuEra Computing. The full paper is on arXiv. That author list matters: this is not a demo staged for a launch event, it is a research result with named domain experts attached, published in a form other specialists can pull apart.
It is also, importantly, a result about paper designs for future machines — not a quantum computer running Claude’s circuits today. Hold both halves of that sentence at once. The designs are real, checkable and, in places, strikingly better than what humans had published. And no useful quantum computer exists yet to run them at scale.
How the loop worked
The setup is the part every builder should steal, because it explains why this task worked when so many autonomous-research claims do not.
- Claude wrote each circuit as a Python program, starting from a plain-language specification of what the circuit had to compute.
- An automated evaluator checked correctness first, testing candidates against large numbers of inputs, and only then scored them.
- The score was resource cost: roughly, the number of qubits multiplied by the number of gates — a proxy for how expensive the circuit would be to run on a real machine, where qubits are scarce and every gate is a chance for error.
- The score went back to a fresh round. Claude revised, tried different mathematical structures, and kept going with no human edits between rounds. Some of the underlying algebraic claims were additionally formalised in Lean, the proof assistant, so they could be verified rather than merely believed.
Quantum circuit design is almost perfectly shaped for this. Circuits can be written as code. Correctness, for the class of algorithms involved, can be checked efficiently and objectively. There is a cheap, trustworthy scorer, and the model can attempt thousands of variations against it. That is the same pattern behind the strongest recent results in AI-assisted mathematics: progress comes fastest where the answer can be graded by something other than another model’s opinion.
The two headline results
Across the eight tasks — spanning chemistry, optimisation, physics simulation, number theory and numerical computing — Claude’s designs reduced both gate counts and qubit counts in seven. The eighth produced a new reversible architecture for modular inversion, a building block where the comparison is less direct. Two results stand out.
17 times fewer gates for elliptic-curve point addition
Elliptic-curve point addition is a core routine in quantum approaches to breaking elliptic-curve cryptography, so its cost feeds directly into estimates of when such attacks might become practical. From a single-paragraph brief, Claude worked for 53 hours and explored several approaches in parallel.
The obstacle was modular inversion. A classical technique called divsteps makes branching decisions at each step, and a reversible quantum version normally has to store a long record of those decisions — which costs many extra qubits. Claude’s key move was to derive algebraic identities that let the branch history be decoded from the numbers themselves after each iteration, so the record never has to be stored at all. The identities were checked with large-scale simulations and formalised in Lean.
The resulting family of designs expands the trade-off frontier between the two scarce resources. The headline point on it uses about 12 percent more qubits than the smallest previous design — and 17 times fewer gates. That trade-off is the honest version of the headline: not everything got cheaper at once, but the overall cost curve moved a long way.
About 12 times fewer gates for a quantum optimisation decoder
The second result concerns decoded quantum interferometry, or DQI, a candidate route to a large quantum speedup on optimisation problems. Its bottleneck is a Reed-Solomon decoder. Google’s earlier construction used the extended Euclidean algorithm, whose input-dependent steps sit awkwardly in a fixed sequence of quantum gates.
Claude replaced it with Berlekamp-Massey decoding, which runs the same steps every round. The catch was storage: a naive reversible version would need to save intermediate records that could add up to hundreds of extra qubits. Claude’s design reconstructs most temporary values from the circuit’s current state instead, and used simulation — some 403 million sample decodings — to show that a single saved value of 19 qubits was enough where a worst-case design might have needed 665. The price is a small, acknowledged failure probability of about one run in 10,000, which the researchers say sits within the task’s permitted error rate. Together, the changes cut the decoder’s gate count by roughly 12 times against Google’s prior construction.
Why gate counts are worth fighting over
It is tempting to read “17x fewer gates” as an abstract benchmark. In quantum computing it is closer to a timeline variable. Anthropic’s own illustration: a 2017 estimate put the cost of computing the ground-state energy of FeMoco — a molecule central to nitrogen fixation — at about 50 trillion Toffoli gates. Later algorithmic advances brought that down to roughly 340 million, a reduction of about 150,000 times, corresponding to a few hours on a future machine with around three million physical qubits.
Those gains did not come from better hardware. They came from smarter circuits: for years, the fastest progress toward a useful quantum computer has been made by reducing what the machine is asked to do. A tool that can grind through that reduction work autonomously attacks the bottleneck that has actually been moving.
What this does not prove
Three limits deserve the same prominence as the results:
- These are resource counts, not hardware runs. Error correction, qubit quality, control systems and connectivity can all change what a paper-efficient circuit costs in practice. None of that is addressed here.
- Independent replication has not happened yet. The paper and baselines are public, so specialists can test the claims directly — but as of publication, the numbers are Anthropic’s, checked by its evaluator and its collaborators.
- The model needed steering at the edges. Anthropic describes Claude as conservative in the loop: it double- and triple-checked apparent improvements, sometimes productively, and it preferred familiar approaches enough that the researchers occasionally had to push it to broaden its search and to try several qualitatively different strategies. Autonomous, in this result, means no human between rounds — not no humans around the work.
There is also a security reading that should not be skipped. Cheaper elliptic-curve circuits lower, on paper, the estimated resources a future fault-tolerant machine would need to attack today’s elliptic-curve cryptography. That does not break any encryption today — the machines do not exist — but it is one more reason security teams should treat resource estimates as moving targets and keep post-quantum migration on a planned schedule rather than a panicked one.
What builders and buyers should take from this
- Invest in the checker before the agent. A 53-hour unsupervised run was only responsible because the evaluator was trustworthy. In enterprise terms: if you cannot score an agent’s output automatically and objectively, you do not have an autonomous workflow — you have an unsupervised one.
- Pick problems with verifiable outputs. Code that must pass tests, designs that must satisfy a simulator, claims that can be formalised — these are where long agent runs compound. Open-ended judgement tasks are where they drift.
- Read AI-for-science claims by their trade-offs. The credible version includes the 12 percent qubit increase and the one-in-10,000 failure rate. Results reported without a price are marketing; results reported with one are engineering.
The bottom line
Nothing about Friday’s paper makes a useful quantum computer arrive this year. What it does show — with named collaborators, a public paper, formal proofs for key claims, and its caveats printed alongside its headlines — is a frontier model doing genuine, verifiable optimisation work in one of the hardest design spaces in computing, and beating the published human record on most tasks it was given.
The pattern is the story as much as the result. Give a model a problem that can be written as code, scored by a machine, and checked by a proof, and it can work for 53 hours and come back with something new. The labs that build those loops — evaluators first, agents second — are the ones whose science claims will keep surviving contact with replication.