Anthropic says Claude Mythos 5 designed circuits for eight computational tasks, and the linked paper’s abstract says seven significantly reduced both Toffoli-gate and qubit counts. The eighth introduced a reversible architecture for modular inversion that improved the trade-off between those resources. Anthropic published its account on October 16, 2026; the paper’s abstract gives the more precise scope of the results (Anthropic; paper abstract).
What the eight-task result says
The result was not a reduction in both measures for every task. The paper’s abstract says seven of eight significantly reduced both Toffoli-gate and qubit counts; the eighth developed a new reversible architecture for prime-field modular inversion and improved the Toffoli–qubit trade-off.
A qubit is a unit of quantum information, while a Toffoli gate is a reversible quantum logic operation. The counts matter in different ways: a circuit can use fewer gates but require more qubits, as one of the reported examples illustrates.
How Claude worked with an evaluator
Anthropic describes a loop in which Claude proposed circuit code, an automated evaluator checked whether it was correct and scored its resource use, and the evaluator’s feedback guided further revisions. Anthropic says there was no human input between those rounds. Its resource score multiplied the number of qubits by the number of gates as a proxy for circuit runtime (Anthropic’s account).
For the elliptic-curve work, Anthropic says the process lasted 53 hours, with minimal human input overall: an initial brief and occasional prompts to think more broadly.
Point addition traded qubits for fewer gates
Elliptic-curve point addition relies heavily on modular inversion. Anthropic says Claude developed a reversible divsteps approach that reconstructs branch information from numerical pairs rather than retaining the full history. The company reports testing candidate rules at 256-bit scale, running numerical simulations and formalizing the result in Lean.
In one comparison for point addition, the design used 12% more qubits but 17 times fewer gates than the comparison design, according to Anthropic. That is a resource trade-off—not a reduction in both measures for that particular design comparison.
The DQI decoder used fewer gates
For decoded quantum interferometry (DQI), Anthropic says Claude replaced the extended Euclidean algorithm in a Reed–Solomon decoder with Berlekamp–Massey decoding, and reconstructed most temporary values instead of storing them. Anthropic reports that the decoder used 12 times fewer gates than Google’s previous construction. This figure is a gate-count comparison for the decoder, not a measured end-to-end speedup for DQI.
The design reserved 19 qubits for one saved exceptional value. Anthropic says a simulator ran 403 million sample decodings to estimate how often that reserve would be exceeded; about one run in 10,000 needed more and failed, within the task’s stated error allowance.
The findings concern circuit design
These are computationally evaluated circuit-design results, rather than a demonstration of a physical quantum computer. Anthropic presents earlier practical quantum computing as a possibility the work could help advance, and says useful systems also require progress in areas such as error correction, decoders and hardware control.
