◆ THE LAB · Research
Anthropic's Claude Mythos 5 designs cheaper quantum circuits
The News
On October 16, 2026, Anthropic said Claude Mythos 5, working in a loop with an evaluator program that checks and scores each design, found more efficient quantum circuits. In seven of eight tasks, including chemistry and optimization, the accompanying paper reports both fewer Toffoli gates and fewer qubits than the baseline it compares each against.
Chandni Melwani
Founder & Editor
Oct 22, 2026 · 1 MIN READ
Know More
- — Authors: Andi Gu (Harvard, Anthropic STEM Fellows Program), Alek Kemeny (Anthropic), Susanne Yelin (Harvard), Mikhail Lukin (Harvard, QuEra).
- — Score: Toffoli count times peak qubit count, weighting time and space equally “for simplicity,” the paper says.
- — HHL rotation angle: 12 times fewer Toffolis and 2.6 times fewer qubits than a tuned baseline, a 32-fold lower score.
- — Fe2S2 selection step: 3.7 times fewer Toffolis and 9.4% fewer qubits than the sparse-compilation baseline, a 4.1-fold lower score.
In 2017, researchers estimated a quantum computer would need about 50 trillion Toffoli gates, among the costliest to run, to compute FeMoco’s ground-state energy. FeMoco is the iron–sulfur cluster in the enzyme bacteria use to turn nitrogen from the air into ammonia. Largely through redesigned circuits, Anthropic says, the estimate has fallen to about 340 million.
Claude began each search from a hand-written specification and a blank program and worked without substantive human intervention, the paper says; Anthropic occasionally prompted it to broaden its search.
On Fe2S2, another iron–sulfur cluster, Claude noticed the molecule’s symmetries and repeating structure meant some results could be calculated once and reused, Anthropic says. For decoded quantum interferometry (DQI), an optimization algorithm, Claude rebuilt the Reed-Solomon decoder, the most expensive step, around Berlekamp-Massey decoding in place of the extended Euclidean algorithm. Combined with other changes, that cut the decoder’s gates by a factor of 12 relative to Google’s previous construction, Anthropic says.
The Read
The paper set Claude's Fe2S2 task in the more direct sparse method because studying the state-of-the-art one “would require validating networks of continuously parameterized Givens rotations acting on superposed states,” outside the circuit model its evaluator uses. Results from this loop for molecules like Fe2S2 will probably keep the sparse baseline until an evaluator can check such rotations.
Room for Disagreement
That limit may not bind. The paper says the state-of-the-art method's factorization “is not unique, and a circuit-aware choice could make its coefficient data repeated or sparse and therefore cheaper to load,” choices it calls analogous to Fe2S2's. The evaluator already scored such data loading on Fe2S2.
Related
Frequently Asked Questions
Did Claude lower the FeMoco estimate?
No. Anthropic says the 150,000-fold drop in that estimate took many research groups eight years, with each step typically representing months of work. Claude's chemistry task was one step in simulating Fe2S2, a different iron–sulfur cluster, and the paper says the amount of compression it found is specific to that molecule's Hamiltonian, though the preprocessing can search other sparse Pauli tables for the same relations and falls back to conventional costs for that step when it finds none.
How were Claude's circuit designs checked?
An evaluator program, which Claude could reach only through the feedback it returned, ran each design on test inputs and checked that outputs matched the target, required inputs were unchanged, scratch qubits returned to zero and the phase agreed exactly. For large input spaces it samples inputs instead of testing all of them. The Reed-Solomon decoder had a failure budget of at most one in 1,000, and seven of 100,000 tests gave a wrong sign. For one design, Claude also wrote formal proofs in Lean.
What caveats come with the results?
Anthropic says Claude “was conservative by nature” in early experiments, second-guessing apparent improvements and preferring well-known approaches, so its setup required several qualitatively different approaches developed in parallel. It also says “Circuit design is only one part of building a useful quantum computer,” which must catch and fix its own errors as it runs. In the eighth task, a number-theory primitive, the paper reports no single baseline, because earlier designs trade Toffolis against qubits differently.
Sources
- Anthropic, “Designing more efficient quantum circuits with Claude” (October 16, 2026): the FeMoco estimates and the eight years they took, the closed loop with Claude Mythos 5, the occasional prompts, the Fe2S2 symmetry insight, the Reed-Solomon decoder's factor-of-12 gate cut relative to Google's previous construction, and the caveats
- Gu, Kemeny, Yelin and Lukin, “Discovering state-of-the-art quantum circuits with large language models,” arXiv preprint 2610.22367 (submitted October 15, 2026): the seven-of-eight result
- The same paper, full text: the authors' affiliations, the work without substantive human intervention, Table I and Extended Data Table 1, the Toffoli-times-qubits score and its equal weighting, the evaluator's checks and sampling, the Reed-Solomon test results, and the Fe2S2 sparse baseline and the state-of-the-art method outside the framework
- Hero photo: the optical table and optical layout of QuEra's first generation neutral atom quantum computer, by QuEra Computing, Inc., Wikimedia Commons, licensed CC BY 4.0
- Creative Commons Attribution 4.0 International licence, which applies to the hero photo
Chandni Melwani
Chandni Melwani is the founder and editor of New in AI, covering AI agents, M&A, and enterprise adoption. She holds a Master's in Management of Artificial Intelligence from Queen's University and brings a practitioner's perspective from her work in Data and AI leadership.
Follow ↗The Lab Briefing
Get the five-beat briefing in your inbox.
The big story, frontier labs, AI at work, e-commerce and the capital crossing borders, each with the part that tells you why it matters, not just that it happened.
Free · No spam · Unsubscribe in one click
More from The Lab
Models
Moonshot's Kimi K3.1 Pro ranks third on Vals' coding test
Moonshot AI's Kimi K3.1 Pro scores 63.64% on Vals.ai's Terminal-Bench 4.0, third of 57 models, at $4.57 a test; Kimi K3 scores 17.17% at $12.48. Moonshot charges a third less for standard K3.1 Pro than for K3, and it promises the weights by November 4.
Chandni Melwani · China · Oct 29, 2026
Models
Reflection's Beam is faster than every model on our table
Artificial Analysis measures Reflection's open-weight Beam at a median 327 tokens a second across its hosts, faster than any model on New in AI's comparison table, but LiveBench scores its reasoning 80.2, below all nine models on that table. Nebius, Beam's cheapest OpenRouter host and the slowest of the five Artificial Analysis tested, charges $0.25 per million input tokens and $1.00 per million output.
Chandni Melwani · US · Oct 23, 2026
Labs
Anthropic is turning away distillers' paying traffic
Anthropic says early evidence suggests its new defences have cut the data harvested by distillers, who covertly extract Claude's capabilities to replicate them in other models, by around 90 to 99%, and that it has turned away meaningful revenue and trillions of tokens' worth of paying traffic from distillers.
Chandni Melwani · Global · Oct 21, 2026