Anthropic reported on October 16, 2026 that its model Claude Mythos 5 autonomously redesigned eight quantum circuits, and on the tasks it was given, it beat the best previously published designs. That is a specific claim, and the details matter more than the headline. Seven of the eight new designs used fewer gates and fewer qubits than the designs they were measured against. The eighth made a trade. It used far fewer gates, but it needed more qubits to do it.
The work came out of a STEM Fellows experiment led by Andi Gu and mentored by Alek Kemeny, with Harvard collaborators Susanne Yelin and Mikhail Lukin. Lukin is also associated with QuEra. An arXiv preprint was submitted on October 15, a day before Anthropic made its announcement public.
What the model actually did
Start with the plain version. A quantum circuit is a set of instructions. Gates are the operations in those instructions, and qubits are the units the operations act on. Smaller circuits are easier to reason about, easier to simulate, and in principle easier to run. So when researchers compare designs for the same task, gate count and qubit count are two of the first numbers they look at.
Mythos 5 was not handed finished answers and asked to tidy them up. According to Anthropic, it worked from plain-language specifications, in a generator-judge loop. Put simply, the system generated candidate designs, an automated evaluator scored them, and the loop kept going. The results were then checked by simulation and by formal proofs written in Lean, a proof assistant that does not accept hand-waving. If a step does not follow, Lean says so.
One run lasted 53 hours. That number is worth pausing on. This was not a quick autocomplete trick or a single lucky guess. It was a long, grinding search, with a machine checking the work as it went.
Two results that stand out
The sharpest result involved elliptic-curve modular inversion. Anthropic says the model’s design used 17 times fewer gates than the smallest prior design for that task. Read that again. Seventeen times fewer. But the company also reported the cost: the design used about 12 percent more qubits. That is a trade-off, not a pure win, and it is to the researchers’ credit that the announcement says so plainly. In quantum computing, you often pay in one currency to save in another.
The second standout was a Reed-Solomon decoder built for Decoded Quantum Interferometry. Anthropic says it used about 12 times fewer gates than Google’s prior construction. The method changed, too. Instead of relying on the extended Euclidean algorithm, the new approach used Berlekamp-Massey decoding. A different route, and by the reported gate count, a much shorter one.
Readers who followed the recent Willow result on cutting quantum errors will see why circuit size matters in the same conversation. Errors are easier to manage when there is less circuit for them to creep into. Shorter designs do not solve error correction, but they give hardware teams less to correct.
Why a generator-judge loop fits this problem
Circuit design is a good match for this kind of setup, and here is why. The rules are strict. A design either performs the task or it does not, and an evaluator can test that without opinion getting in the way. Formal proof adds a second layer: not just “it worked in simulation,” but “the logic holds, step by step.”
That structure is showing up across serious AI research right now. The new funding from Simons, Anthropic and OpenAI for math and AI work points in the same direction – put models against problems where answers can be checked, and let the checking do the disciplining. Mythos 5 and this circuit project fit that pattern neatly.
The caveat, stated clearly
Now the part that should travel with every retelling of this story. These are circuit-design results on paper. They are not a running quantum computer. No machine executed these designs to solve a real workload during this experiment, and Anthropic’s results do not change what still stands between the field and useful systems: error correction and hardware progress. Both are hard. Both are unfinished.
That caveat does not shrink the result. It locates it. Design work is real work, and a library of smaller, proven circuits is exactly the kind of thing hardware teams will need when machines improve. Paper results, properly checked, are how the field stocks its shelves.
There is also a quieter point here about how the work was validated. Simulation can miss edge cases. Human review can miss errors. Lean proofs are slower and stricter than either. Choosing to validate this way, and saying so in the announcement, gives other researchers something solid to push against. If a proof has a gap, they can find it. If it holds, the design earns its place.
What to watch next
Three things, mainly. First, whether other groups reproduce the gate counts and the proof results from the arXiv preprint. Independent checks are the difference between an announcement and a settled finding. Second, whether the elliptic-curve trade-off – 17 times fewer gates for about 12 percent more qubits – turns out to be a bargain hardware teams actually want as qubit counts grow. Third, whether the generator-judge approach transfers cleanly to other circuit families, or whether these eight tasks happened to suit it especially well.
None of that takes the shine off a strong week for the field. A model spent 53 hours on one run, argued with its own evaluator, submitted to formal proof, and came out with designs that beat the published record on eight tasks. Careful claim, careful checking, real numbers. That is how progress is supposed to look.
Be First to Comment