ings
AI has another notable contribution to scientific research.
On October 16, 2026, Anthropic published research showing that Claude Mythos 5 could use natural-language task specifications and feedback from an automated evaluator to improve on previously published quantum circuit designs. The accompanying paper was submitted to arXiv on October 15. Source
Two results stand out: an elliptic-curve point-addition design reduced the gate count to roughly one-seventeenth of its reference design, while a decoder used in a quantum optimization algorithm reduced its gate count to about one-twelfth of the previous construction.
More significantly, Claude derived new algebraic identities during its search and used them to develop and verify a new circuit design.
These results demonstrate a concrete role for AI in research: developing methods, implementing them, and testing the results through mathematics and computation. Reductions in gate count represent improvements in circuit resource requirements; they do not directly establish equivalent speedups on a real quantum computer. Source
1. Quantum Computing’s Bottleneck: Every Qubit and Every Gate Counts
Quantum computing offers new ways to tackle certain problems that are difficult for classical computers, including molecular simulation, materials research, and specific mathematical tasks.
Turning that potential into reality requires overcoming two constraints: high-quality qubits are limited, and quantum states are vulnerable to noise. More complex computations generally place greater demands on hardware and error correction.
Researchers therefore work to reduce the resources algorithms need, allowing the same computation to use fewer qubits and quantum gates.
The difficulty is that these quantum circuits cannot simply adopt the design methods of ordinary software.
First, conditional branches are harder to handle. An ordinary program can choose an execution path based on its input. A coherent quantum circuit acting on a superposition must accommodate the relevant operations within a common circuit structure.
Second, intermediate information cannot simply be erased. In these reversible computations, temporary data must be cleaned up carefully, typically by “uncomputing” the operations that produced it. This adds more gates.
Circuit design becomes a difficult trade-off: retaining intermediate results can avoid repeated computation but requires more qubits, while reducing storage may require more operations.
FeMoco simulation offers a striking example. According to Anthropic’s account, resource estimates for the same 54-orbital model fell from approximately 50 trillion Toffoli gates in 2017 to roughly 340 million in subsequent research, reflecting years of accumulated algorithmic advances. Source
That history shows how profoundly circuit design can change quantum computing’s resource requirements. AI is now beginning to contribute to this effort.
2. A 53-Hour Search: Putting the Model Inside a Design-and-Verification Loop
Through Anthropic’s STEM Fellows Program, researchers selected eight foundational quantum-circuit tasks spanning chemistry, optimization, physics simulation, and number theory.
The researchers supplied Claude with natural-language specifications and an automated evaluator. Claude implemented designs in Python, the evaluator checked them and calculated resource costs, and the results were returned to the model to guide the next revision.
One central scoring metric was the product of qubit count and gate count. This gave the model a clear objective: satisfy the task requirements while finding implementations with lower resource costs.
According to the paper, seven of the eight tasks achieved significant reductions in both Toffoli gate counts and qubit counts. The eighth introduced a new reversible modular-inversion architecture that improved the trade-off between those resources. Source
The widely discussed elliptic-curve result emerged from a 53-hour search. Anthropic describes the human input as primarily the initial brief and occasional prompts to broaden the search. The model then developed and verified its approach. Source
Those 53 hours refer to that specific task. They should not be interpreted as reproducing eight years of FeMoco research in just over two days.
Its significance lies in showing that, within a suitable evaluation environment, a model can explore complex solutions and turn new ideas into implementations that can be tested.
3. Breakthrough One: Recovering the “Scratch Work” from the Result
The first breakthrough concerns elliptic-curve point addition, where modular inversion is a major source of computational cost.
Informally, modular inversion means finding a number’s reciprocal under modular arithmetic. It is a basic operation that appears repeatedly in the relevant quantum algorithms.
In classical computing, a method called divsteps handles this efficiently through repeated swaps, additions, subtractions, and halving operations.
Converting it into a reversible quantum circuit creates a problem: the operation selected at each iteration generally needs to be recorded. More iterations mean more space devoted to storing that branch history.
Claude explored a different possibility: could the previous branch choice be recovered directly from the updated numerical state?
If that were possible, storing the complete history separately would become unnecessary.
Early rules worked on small examples but failed at the full 256-bit problem size. Claude then investigated the mathematical structure of the iteration process, derived new algebraic identities, and used them to construct a branch-recovery rule.
According to Anthropic, Claude followed this with large-scale numerical checks and formalized the relevant claim in Lean, producing a reversible divsteps implementation that requires no stored branch history. Source
It then explored different combinations of storing and reconstructing information, producing a family of designs with different resource trade-offs.
One design used about 12% more qubits than the smallest previous design, while requiring roughly one-seventeenth as many gates. Source
This gave researchers a previously unavailable option: a modest increase in space in exchange for substantially fewer computational operations.
4. Breakthrough Two: Changing the Decoder to Cut the Gate Count Twelvefold
The second breakthrough concerns decoded quantum interferometry, or DQI.
In the task examined by the researchers, Reed–Solomon decoding was a major source of cost. The previous circuit used the extended Euclidean algorithm, whose execution depends heavily on the input.
An ordinary computer can adapt its execution path, whereas the quantum circuit must accommodate the relevant operations within a fixed structure, introducing additional overhead.
Claude switched to Berlekamp–Massey decoding. Its more regular iteration structure makes it better suited to a fixed quantum circuit.
Changing methods was only the first step. The new reversible implementation still generated temporary data, and retaining all of it would remain expensive.
Claude again used reconstruction, recovering most intermediate information from the current state and reserving storage only for exceptional values that could not be recovered.
It then wrote a high-speed simulator and ran approximately 403 million sample decodings to measure how frequently those exceptions occurred.
The results showed that almost every run needed to retain at most one exceptional value. The storage budget for those exceptions was therefore reduced from 665 qubits in the worst case to 19. These figures do not represent the decoder’s total qubit count.
That trade-off had an explicit cost: roughly one run in 10,000 would fail because it required more storage, remaining within the task’s permitted error rate. Source
Changing the algorithm, reconstructing temporary information, and sizing resources using measured statistics together reduced the decoder’s gate count to approximately one-twelfth of the previous design. Source
The speedup for the complete task on a real machine would still depend on other computational steps, error-correction overhead, and hardware.
5. A Shift in Research: AI Is Beginning to Develop New Methods
The research raises a broader question: how might scientific work change as AI becomes capable of contributing to method development and mathematical reasoning?
People are already familiar with AI summarizing papers, explaining concepts, and assisting with programming. This research demonstrates its potential to contribute to algorithmic exploration: changing direction in response to feedback, proposing structural improvements, and checking whether they work.
These results suggest that, in tasks with clear objectives and testable outcomes, models may take on more iterative exploration, implementation, and verification.
Researchers still need to choose meaningful problems, establish evaluation criteria, examine assumptions, and determine where the conclusions apply.
A productive question is therefore how to combine a model’s exploratory capabilities with researchers’ expertise to produce reliable results more efficiently.
6. What Comes Next: From Better Circuits to Joint Design Across Layers
This research demonstrates a way to reduce the resources quantum computations require. Hardware quality, error correction, and systems engineering still require further progress.
When algorithms require fewer gates and less storage, some applications may move closer to the capabilities of future hardware. Advances in circuit design can work alongside improvements in hardware and error correction.
In its discussion, Anthropic suggests that future AI systems could jointly optimize algorithms, circuits, error correction, and hardware constraints to find better overall designs for specific applications. This remains a direction for further research. Source
Such designs could also help hardware teams identify which capabilities deserve priority when building machines for a valuable application.
The 53-hour search offers a research model worth testing further: people define the problem and standards, AI develops candidate solutions, and computation and mathematics test the results.
If that cycle can work reliably across more tasks, AI’s role in scientific discovery could expand further.
Cover image credit: Anthropic.