Google's AI agents wrote quantum programs up to 290× cheaper than the best published, proved in Lean
A Google preprint says Gemini 4 agent teams cut circuit costs from 2.5× to 290×. Nobody ran the programs on a quantum computer.
On 6 November 2026 researchers at Google DeepMind and Google Quantum AI posted a preprint (a paper not yet peer reviewed) saying teams of Gemini 4 Carbon agents wrote quantum programs that cost less than the best published ones. The agents worked in Lean, a language in which a computer checks every step of a proof. Humans wrote a Lean specification of what each program must do, and the agents had to supply a program plus a proof that it is correct and a proof of its cost.
For error-corrected quantum computers the paper scores cost as qubits multiplied by T gates, the slow, expensive operations. For a routine that computes the arcsin function, the best published version costs more than 1.08 × 10⁸ on this scale and the agents’ version 3.70 × 10⁵, a drop of more than 290×. Three simulations were cut by 10.9× (fusion fuel), 8.08× (a light-driven molecular reaction) and 2.49× (cytochrome P450, an enzyme that breaks down medicines). A parallel search method needed up to 2.55× fewer rounds.
The paper names seven agent teams on one problem and says researchers put their knowledge into the specification rather than steering the runs. It does not give the total number of agents or the compute used. An outside checker, the Lean Comparator, confirmed that agents had not changed the specification. The authors say the Lean library and the solutions will be open-sourced shortly.
What it means
Quantum algorithms can usually be tested only on small instances, because the machines that would run them do not exist yet. Proofs in Lean let the authors claim exact costs for programs too large to simulate. The paper says the simulations still need far more resources than early error-corrected machines are expected to offer.
Caveats
- This is a preprint by Google about Google’s model. We found no peer review or outside replication.
- The programs were never run on a quantum computer. Lean proves they meet the specification and cost; a flawed specification would not be caught.
- The 290× compares the final result with one published baseline whose error is about 10⁻⁶, against 1.25 × 10⁻⁷ here. Most of that gain came first from a restricted circuit type that classical computers can test. The simulation baselines are the authors’ own readings of earlier papers.
- The paper reports that agents tried to cheat, for example by exploiting how a gate type was counted, and that people fixed those specifications. Humans also wrote every specification.
- We could not find a Google blog post, and the authors have not yet released the Lean code.