Claude sets records in 7 quantum circuit tasks

Summary

A team from Harvard and Anthropic used Claude with an evolutionary search method (OpenEvolve) to redesign quantum circuits. They improved the total Toffoli gate × qubit cost in seven out of eight tasks. The claim of a '17-fold' reduction in 256-bit elliptic curve point addition is real; however, this measurement was made against a baseline using fewer qubits, not the overall best design.

Why it matters

Quantum algorithm costs have historically been reduced more by better circuit design than by hardware advancements. Therefore, an AI that independently proposes, implements, and verifies circuit architectures can shorten the time required for useful quantum computing without touching a single physical qubit. The model proposes, the classical oracle verifies, and the evolutionary loop repeats without human intervention. This is a general workflow that extends far beyond quantum computing.

Highlights

• The 17-fold figure is real but selective. A competing 2026 design (Schrottenloher) outperforms Claude in total cost for the same elliptic curve task. This is why the improvement factor for that task is not listed in the paper's main table.
• The 53 hours of fully autonomous work on the elliptic curve task required only the initial prompt and occasional 'think broader' instructions. The correctness of one circuit was also formalized using the Lean proof assistant.
• Anthropic's blog post claims improvements across all eight tasks. The 46-page arXiv paper, however, notes improvements in seven out of eight. The paper is more conservative and likely the more accurate source.

Critical take

The result is a preprint that has not undergone peer review or independent replication. The authors themselves acknowledge that Claude is 'conservative by nature' and had to be forced to explore parallel discoveries. Therefore, it is unclear how much of the gain comes from the model versus the evolutionary search framework. The total cost metric is explicitly arbitrary. The 17-fold claim hides qubit overhead. The elliptic curve in question is the one used by Bitcoin. Thus, the 'no threat' framing is not a neutral observation but a practical leap.