Claude Designs Leaner Quantum Circuits, With a Computer Checking Every Step
Anthropic says Claude Mythos 5 found quantum circuit designs that need fewer qubits and gates than published ones across eight tasks. A built-in checker kept it honest.
Anthropic reports that Claude Mythos 5 designed quantum circuits that beat the best published designs on eight tasks, from chemistry to number theory. Some improvements were, in Anthropic’s words, “an order of magnitude or more”.
A quick translation first. A quantum circuit is the recipe a quantum computer follows, built from qubits (the quantum version of bits) and gates (the individual steps). Fewer qubits and fewer gates means a smaller, cheaper machine could run the same job. That matters because today’s quantum hardware is small and error-prone.
Here is how it worked. Claude got plain-language descriptions of each task and wrote circuit designs in Python. An automated evaluator checked that each design was correct and scored it by qubits times gates. Claude then revised its work from that feedback in a loop, with no human in between rounds. Two examples from the post: for a step used in elliptic-curve calculations, one design uses 12 percent more qubits but 17 times fewer gates than the smallest earlier version, after a 53-hour run. For a decoder used in a quantum optimisation method, Claude switched to a different algorithm and reused temporary values, reaching about 12 times fewer gates than Google’s earlier construction.
What’s behind it: this works because the answers can be checked cheaply on an ordinary computer. The evaluator, large numerical simulations (403 million sample decodings for one decoder) and a formal proof in Lean, a proof checking tool, mean you do not have to simply trust the AI. Anthropic says this kind of problem is “fairly common” among the algorithms it studied. It also reports that Claude was cautious, over-checked itself and leaned toward familiar methods, so the team asked for several different approaches in parallel.
Keep it in proportion. The post does not say the results are peer reviewed (a paper is on arXiv, 2610.22367). Circuit design is only one part of building a useful quantum computer, and error correction and hardware are separate hard problems. The famous 150,000-fold drop in estimated gates for one chemistry problem came from many groups over eight years, not from Claude.
What this means for you: no quantum computer is arriving on your desk. The lesson is about method: AI works best on problems where a machine can verify the answer, and there it can produce real, checkable progress. That pattern is worth watching in science and engineering.
Sources
Source: https://www.anthropic.com/research/designing-efficient-quantum-circuits
ChatGPT Pro Can Now Run Several Agents on One Task at the Same Time
OpenAI says Pro users can split a job across parallel agents in ChatGPT and watch their progress. Details are thin, so here is what is known and what to wait for.
