Anthropic says its Claude Mythos 5 model has designed quantum circuits that improve on the best previously published designs across a set of eight basic computational routines, in some cases by a wide margin. Why it matters: circuit design is one of the bottlenecks holding back useful quantum computing, and this is a claim that an AI model can do original work in that field largely on its own, not just assist a human expert.
October 16, 2026 — Anthropic published the research on Friday, alongside a companion paper submitted to arXiv on October 15.
What happened
According to Anthropic, Claude worked in a closed loop with an automated evaluator. The model proposed circuit designs in Python, the evaluator checked each design for correctness and scored it on the resources it would need, and Claude revised its work based on that feedback, with no human input between rounds.
The eight tasks covered building-block routines used in larger quantum algorithms, spanning areas such as chemistry, optimization, quantum simulation and number theory. Anthropic says the new designs reduced both gate counts and qubit counts compared with the prior state of the art in seven of the eight tasks, and that the eighth produced a new reversible design for a number-theory routine.
Two results Anthropic highlights
The first concerns a step used in elliptic-curve arithmetic. Anthropic says Claude worked on it for 53 hours with minimal prompting and produced the first reversible version of a classical method that avoids storing a record of every branch taken along the way — a storage problem that had made the method impractical for quantum circuits. One design in the resulting family uses somewhat more qubits than the smallest previous design but far fewer gates.
The second concerns a decoder used in a quantum optimization approach. Anthropic says Claude reorganized the circuit around a different decoding method and used simulation to size its memory more tightly, cutting the decoder’s gate count by a factor of 12 compared with the earlier construction it measured against.
Why it matters
Quantum circuits are expensive to design well because every extra gate and qubit raises the cost of running them on future hardware. If AI models can reliably find leaner designs, development timelines for quantum algorithms could shorten. For Tested Intelligence readers, the subscriber take is simpler: this is another data point that frontier models are moving from writing text and code into producing verifiable scientific work, where an independent checker can score the output rather than a human judging style.
The caution matters just as much. These are Anthropic’s own reported results in a preprint that has not yet been independently validated or tested on quantum hardware. Treat the headline numbers as a strong claim awaiting a second opinion, not a settled result.
Frequently asked questions
Did Claude design these circuits by itself?
Anthropic says the model worked largely autonomously inside an automated loop, proposing designs and revising them from evaluator feedback, with humans setting up the tasks and occasionally prompting it to broaden its search.
How much better were the new circuits?
Anthropic reports improvements across all eight tasks, with seven reducing both gate and qubit counts against prior published designs. The largest single figures it highlights are a 17-fold reduction in gates for one design family and a 12-fold reduction for a decoder circuit.
Has anyone verified the results?
Not independently yet. The work is described in an arXiv preprint submitted October 15 by the Anthropic and Harvard collaborators, and Anthropic says the model checked one key result with formal proofs and large-scale simulations.
Sources: Anthropic research post, arXiv preprint, Unite.AI, Crypto Briefing