Anthropic published STEM Fellows research on Oct 16: given plain-language specs plus an automated evaluator that verifies correctness and scores qubits x gates, Claude Mythos 5 found quantum-circuit designs that beat the best published ones across eight primitives in chemistry, optimization, simulation and number theory. On elliptic-curve point addition it worked 53 hours with minimal human input and produced the first reversible divsteps modular inversion with no stored branch history; for decoded quantum interferometry it cut Reed-Solomon decoder gates about 12x versus Google's prior construction. Paper: arXiv 2610.22367.

Key Takeaways

  • ✓Beat the best published designs on all 8 quantum-circuit primitives; the paper says 7 of 8 cut both Toffoli and qubit counts (arXiv 2610.22367)
  • ✓Point addition: 53 hours of largely autonomous work yielded the first reversible divsteps modular inversion with no stored branch history, formalized in Lean
  • ✓One design in the new family uses 12% more qubits but 17x fewer gates than the smallest previous design
  • ✓DQI Reed-Solomon decoder: ~12x fewer gates than Google's prior construction; 403M sampled decodings shrank a 665-qubit worst-case reserve to 19 qubits (~1-in-10,000 failure)
  • ✓Reusable pattern: LLM writes Python circuits, an independent evaluator verifies and scores qubits x gates, and the loop runs without human input between rounds
Anthropic: Claude Mythos 5 designs quantum circuits in an evaluator loop, beating best published designs on 8 tasks and cutting a DQI decoder's gates 12x
🖼️Official Media
Click to view high-res
🧭

Turn your technical choice into a development budget

Compare 40 dev plans & simulate token costs vs $20/mo subscriptions

🔬

In-Depth Technical Analysis

Anthropic's STEM Fellows team (led by Andi Gu, with Harvard's Susanne Yelin and Mikhail Lukin) gave Claude Mythos 5 plain-language specs for eight quantum-circuit primitives plus an automated evaluator that checks correctness and scores qubits x gates. Working in a closed loop with no human input between rounds, Claude beat the best published designs on all eight; the arXiv paper (2610.22367) reports that seven cut both Toffoli and qubit counts. On elliptic-curve point addition, Claude worked 53 hours, proved new algebraic identities and produced the first reversible divsteps modular inversion that stores no branch history, verified in Lean; one design uses 12% more qubits but 17x fewer gates than the smallest prior design. For decoded quantum interferometry, switching to Berlekamp-Massey and reconstructing temporaries cut the Reed-Solomon decoder's gates ~12x versus Google's construction, using 403M sampled decodings to shrink a 665-qubit worst-case reserve to 19 qubits. All figures are Anthropic-reported. The takeaway for developers is the evaluator-guided loop: any problem with machine-checkable outputs and a scalar objective can be handed to a coding agent for long autonomous optimization.

Action HubReady to adopt this in production?

Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.

Up NextClaude traces the NIST elliptic-curve seeds to lrand48 seeded with 1997–98 timestamps, settling a 25-year backdoor question
Read Now→