Anthropic says Claude cut storage in a quantum decoder

In an account published October 16, 2026, Anthropic said Claude redesigned a Reed–Solomon decoder for decoded quantum interferometry (DQI) to reserve 19 qubits instead of 665 for a worst-case design. The approach reconstructed most temporary values rather than storing them; in Anthropic’s simulations, about one in 10,000 sample decodings needed more storage than the design reserved, a rate the company said was within the task’s allowed error rate. Anthropic’s research account

How Claude redesigned the decoder

A reversible quantum circuit must be designed so its operations can be undone. Anthropic said the decoder’s earlier approach used the extended Euclidean algorithm, whose input-dependent sequence of operations is awkward to implement as a fixed reversible circuit. Claude instead used Berlekamp–Massey decoding and reconstructed most temporary values from the circuit’s current state, rather than keeping those values in storage.

Anthropic reported that the redesigned decoder used 12 times fewer gates than the previous construction. The reduction in stored qubits came from a related but distinct choice: reserve space for one exceptional saved value, rather than enough space for the worst case across the process.

What the 19-qubit reserve means

Anthropic said a worst-case reserve would require 665 qubits, while Claude’s design reserved 19. In 403 million simulated sample decodings, about one in 10,000 needed more storage than that reserve provided. Anthropic said the overflow rate was within the task’s allowed error rate.

So 19 qubits was a deliberate storage trade-off, not a guarantee that every possible decoding would fit. The overflow figure describes Anthropic’s simulation of this decoder task; it is not a measurement of a quantum processor running the circuit.

Where the decoder fits in Claude’s quantum-circuit research

The decoder was one of eight computational tasks in a preprint submitted to arXiv on October 15, 2026. Its abstract reports significant reductions in both Toffoli-gate and qubit counts in seven of the eight tasks. For the remaining task, modular inversion, the authors describe a new reversible architecture that improved the trade-off between those resources. The arXiv preprint

Anthropic described Claude Mythos 5 proposing circuit code, receiving correctness feedback and resource scores from an evaluator, then revising its designs. The findings concern circuit design and simulated evaluation—not a quantum processor executing the circuits. Anthropic reported the decoder’s resource savings as part of that design work.