Better quantum circuit design is one of the quieter ways to make quantum computers useful sooner, and Anthropic says its Claude model just did a lot of it. In a new research post, the company reports that a model it names Claude Mythos 5 produced circuits that beat the best published designs on all eight tasks it tried, including a roughly 12x cut in the gate count of one expensive step.
The results are self-reported and come with a long list of caveats the authors state themselves. Here is what the post claims, how it was checked, and what it leaves untouched.
In this article
- What Anthropic says Claude did
- The numbers behind the quantum circuit design claims
- How the results were checked
- Why the self-checking loop is the real story
- What this does not change
- Editor’s take
What Anthropic says Claude did
According to Anthropic’s research page, Claude designed circuits for eight computational building blocks spanning chemistry, optimization, quantum simulation, number theory and numerical computing. Two are highlighted: elliptic-curve point addition, and the decoder inside Decoded Quantum Interferometry, an optimization algorithm known as DQI.
The setup matters for how much weight to give it. Claude received plain-language specifications and an automated evaluator, and it worked in a closed loop with no human input between rounds. The evaluator checks that a proposed circuit is correct and scores it by multiplying qubits by gates, so a design only wins if it is both right and lean.
The work was led by Andi Gu as part of Anthropic’s STEM Fellows Program, with mentorship from Alek Kemeny. Collaborators included Susanne Yelin and Mikhail Lukin of Harvard, with Lukin also at QuEra Computing. The post names real outside collaborators, which is a good sign for scrutiny, though it does not say whether the work has been peer reviewed. A full paper is linked on arXiv.
The numbers behind the quantum circuit design claims
Anthropic says Claude’s designs used fewer qubits and fewer gates than the best published circuits for all eight tasks, and that some improvements were an order of magnitude or more. Two figures stand out.
The first is the 12x result. The most expensive step of DQI, a Reed-Solomon decoder, has about 12 times fewer gates than Google’s previous construction. The second is a trade-off rather than a pure win: one new point-addition design uses 12% more qubits than the smallest earlier design but 17 times fewer gates. That run took 53 hours.
That trade-off is worth pausing on, because quantum circuit design is always a balancing act. Qubits are scarce, gates add errors and runtime, and the best answer depends on which constraint binds on a given machine. A design that spends a few more qubits to save a large number of gates may be excellent for one machine and unhelpful for another. The post’s scoring rule, qubits times gates, is a reasonable single number, but it is a choice.
How the results were checked
The evaluator verifies each circuit’s correctness as part of the loop. For one design, a modular inversion method called divsteps, Anthropic says it also ran large-scale numerical simulations and formalized the result in Lean, a proof assistant. For the decoder’s memory approach, the team simulated 403 million sample decodings and reports failure in about one run in 10,000, which it describes as within the allowed error rate.
Those are meaningful checks, because they do not depend on trusting the model. A circuit either computes the right answer or it does not, and a proof assistant does not care who wrote the design. Still, the post is a company’s own account, so independent replication by other groups is the real test.
Why the self-checking loop is the real story
Strip away the quantum vocabulary and the pattern is simple: a model proposes, a fast automatic judge scores, and the loop repeats for hours. That works because quantum circuit design has an unusually clean referee. Correctness can be tested mechanically, and the score is a single number, so the model cannot talk its way past a bad design.
Most real-world work is not like that. Writing, strategy and policy have no automatic evaluator, which is why results like this one should not be read as proof that models can do any hard job unsupervised. They are better read as evidence that when a task has a trustworthy checker, a model can search a design space faster than a human team can.
The authors also say the model’s habits shaped the process. It was cautious, it second-guessed itself and it leaned toward familiar methods, so the team had to push it to explore. That is useful context for anyone planning to point a model at a similar problem: the tool needs steering, and the steering is part of the result.
What this does not change
The authors are direct about limits. Claude was conservative: it repeatedly second-guessed results and tended to prefer well-known approaches. The team countered that by requiring parallel approaches and occasionally prompting the model to broaden its search.
More important, this is circuit design only. Practical quantum computing also needs error correction, fast decoders and hardware control, and nothing here solves those. A leaner circuit still has to run on machines that make mistakes.
One headline number needs care. The post mentions a roughly 150,000-fold resource reduction for a chemistry problem known as FeMoco, but it says that came from many groups’ algorithmic advances over several years, not from Claude alone. And the claim that practical quantum computers could arrive sooner is a forward-looking expectation, not a demonstrated outcome.
Editor’s take
Watch the replication, not the multiplier. A 12x figure is eye-catching, but the credible version of this story is a model that can do a well-scoped optimization task with an automatic checker, and a team that published how it checked the answers. That is a template other fields can copy: give the model a fast, trustworthy evaluator and let it grind.
The risk is reading it as a quantum breakthrough. Gate counts are one input to a hardware race that depends on error rates, scale and engineering. For related reading on how AI labs report their own results, see our coverage of OpenAI’s disclosure of four more agent incidents, and for the business angle on AI claims, what the new AI exclusions mean for your business.
What to watch next: outside groups reproducing the quantum circuit design results, and whether the savings hold up on real hardware constraints.
