Frontier Signal · Deep Dive · AI & Quantum · 7 min read · Monday, 19 October 2026
On 15 October, researchers at Harvard and Anthropic reported that Claude designed quantum circuits with far fewer steps than comparison designs on seven of eight tasks. A circuit is the recipe a quantum computer follows, and every step costs a future machine time and hardware. Imagine a recipe that drops from 5.1 million steps to 440,000, with slightly fewer kitchen counters needed. The big questions are who checked it, what it was compared with, and whether the kitchen can be built. Here, the steps are quantum gates and the counters are qubits.
That is the headline from the preprint, a paper shared before other scientists have reviewed it. An AI model wrote the quantum circuits and an automatic checker tested each one. Qubits, as our review of the year error correction went commercial explained, are the scarce part.
12× LOWER COST ON ONE KEY STEP · 8 TASKS STUDIED · 0 RUNS ON REAL HARDWARE
Cost means the number of basic steps (Toffoli gates) multiplied by the peak number of qubits, from the paper’s Table 1. These are simulations and estimates. The paper has not been peer-reviewed.
Key takeaways
Claude designed circuits for eight quantum building blocks. It worked in a loop that tries ideas and keeps the best, with an automatic checker. Seven improved on both steps and qubits against their comparison designs. One comparison was a basic version the researchers built themselves.
Headline case: a decoder step (called DQI) fell from about 5.1 million to about 440,000 steps, roughly 12 times fewer. Total qubits moved from 1,938 to 1,837.
The strongest comparisons against published designs show about 4 to 39 times lower cost. A gain in one small part does not give the same gain for a whole quantum application.
This is not a demonstration of breaking Bitcoin. The work does not give a new full estimate for such an attack.
No hardware test or peer review yet. The authors’ ties to Anthropic and QuEra make independent checking especially valuable.
Here is what happened
On 15 October 2026, Andi Gu, Alek Kemeny, Susanne Yelin and Mikhail Lukin posted a paper called “Discovering state-of-the-art quantum circuits with large language models” (arXiv 2610.22367).
Kemeny works at Anthropic and Gu had Anthropic fellowship support. Yelin and Lukin are at Harvard, and Lukin co-founded the quantum company QuEra.
Anthropic posted its own account on 16 October.
The task: design reversible circuits for eight building blocks, scored by steps multiplied by qubits. From the paper’s Table I (approximate):
DQI decoder (Reed-Solomon): 5.1 million to 440,000 steps; 1,938 to 1,837 qubits; about 12 times better.
Fixed-point Gaussian function: about 24,000 to 1,583 steps; 362 to 139 qubits; 39 times.
Rotation step for solving linear equations (HHL): about 16,000 to 1,348 steps; 300 to 115 qubits; 32 times.
Three more (a Coulomb potential, an iron-sulphur chemistry step and a lattice-gauge simulation) improved 4 to 10 times.
Hermitian DQI decoder: about 100 million to 930,000 steps, but the comparison is a basic version the authors built because none existed. We do not count it as a fair match.
Prime-field modular inversion: about 4.0 million steps and 935 qubits, with no single comparison in the paper.
Seven tasks improve on both measures against the comparisons in Table 1. Six use published designs. The Hermitian decoder uses the basic version. The eighth task, modular inversion, is judged through a trade-off rather than one headline comparison.
How it works
A human sets the task, not the answer. The team gave Claude a specification and a shared Python toolkit. Claude Mythos 5 ran inside OpenEvolve, an open tool that proposes ideas, tests them and keeps the best, a pattern DeepMind popularised with AlphaEvolve. It starts from a blank program and improves up to five families of approach in parallel.
An automatic checker is the referee. It simulates the allowed circuit operations on test inputs, checks the outputs and counts steps and qubits.
The method suits a limited kind of circuit that classical computers can check efficiently. Passing sampled tests is not proof for every possible input.
Why the researchers multiply steps by qubits
Toffoli gates are the costly steps in error-corrected quantum computing.
Their count stands in for the work done, while the peak number of qubits measures the computer’s size. Multiplying them gives one score to compare designs.
Actual running time depends on more: circuit depth, routing, error correction and the supply of special resource states. The score is not a universal price tag.
Why DQI and elliptic curves? DQI turns an optimisation problem into a decoding problem, and decoding is the costly step.
Elliptic curves protect Bitcoin and Ethereum signatures. Breaking them takes very many point additions, each needing a modular inversion (a kind of division). Shrinking inversion shrinks one block of that attack, not the whole.
THE OTHER WAY TO CUT THE QUBIT BILL
This paper shrinks the algorithm. Another route shrinks the error-correcting layer underneath it: our look at the codes that cut qubit overhead argues the real problem is needing too many physical qubits for each reliable one.
What the headline savings actually measure
Both big wins use the same score. The 12-times (decoder) and 39-times (Gaussian) improvements each compare a specific task, not a complete machine.
The Hermitian decoder’s huge saving is not like for like. Its comparison is a basic version, so do not present it as a gain over the best published design.
Most of the decoder’s gain comes from fewer steps, not fewer qubits. For the Reed-Solomon decoder, qubits fall only from 1,938 to 1,837, about 5.2%. A big drop in one internal stage must not be mistaken for a drop across the whole circuit.
What it does and does not change for encryption
A building block is not a full attack. The modular-inversion row reports 4.0 million steps and 935 qubits for its whole test circuit. That cannot be compared directly with what it takes to break a cryptographic key.
Logical and physical qubits are different. Turning a design into a hardware estimate needs an error-correction plan and assumptions about the machine.
This paper alone does not change the timetable for upgrading encryption. For the wider debate, see our look at Bitcoin’s quantum freeze-or-lose fight.
Who captures the value
Possible users are teams building quantum hardware and software. Better circuits could cut the resources their programs need, but the paper does not show commercial use, revenue or any company’s earnings.
For investors, the useful signal is progress in algorithm design alongside hardware. Build quality and error correction still matter, as the blind study of weak qubits and the error-correction story above show.
Commercial value depends on whether these gains survive when put into full applications.
The honest catch
Estimates, not hardware. Real cost depends on routing, error correction and the supply of special resource states. The steps-times-qubits score is a chosen measure, not a real price.
Baselines matter. The Hermitian decoder uses a basic comparison, and modular inversion has no single comparison in Table 1. Keep those apart from comparisons with published designs.
Some savings allow a small failure rate. In the inversion design, one memory limit overflows in about 2.4 per million sampled runs, and a separate time limit fails about once in 100,000. The first number alone is not the total failure rate.
A proof and a test cover different things. A mathematical argument for one part does not certify the whole circuit. The complete Reed-Solomon circuit produced seven wrong signs in 100,000 reported tests.
Eight chosen tasks are not all of quantum computing. The paper does not give a full account of what the search itself cost.
Outside checking matters. Others should test both the comparisons and the effect of the allowed errors in full algorithms.
The one question every reader should ask
Against what baseline, and what did the checker not test?
The key questions are whether the comparison is fair, which inputs the tests cover, and whether the allowed error stays acceptable once the circuit is part of a bigger calculation.
OUR VIEW
The frontier signal is AI helping to find circuit designs that can be tested and scored. The paper shows large simulated gains on specific tasks. Hardware use and independent checks are still open. It is promising research for people designing error-corrected quantum computers, not evidence that quantum computers can now break encryption.
What to watch next
Independent checks with the released code and checker.
Fair comparisons: published baselines, matched accuracy needs and full application costs.
Reactions from cryptographers and algorithm researchers.
Real-world costs: do the gains survive routing and special-state overheads?
Peer review, and any team that says it will use the circuits.
Quick questions
What did Claude actually do?
It proposed and improved programs that generate circuits, guided by an automatic checker. Seven tasks improved on both steps and qubits against their comparisons. One comparison was a basic version, not earlier published work.
Does this make Bitcoin easier to break?
It improves some building blocks. It does not show an attack or give the full hardware needed.
Does 39 times lower cost mean 39 times faster?
No. It is a cut in steps multiplied by qubits for one task. Running time and real machine cost need more assumptions.
Can a quantum computer run these circuits today?
No. They are designs for large, error-corrected machines that do not exist yet.
Why does the conflict of interest matter?
Anthropic built the model and published the work, one author works there and another co-founded QuEra. That does not make the results wrong. It makes outside checking matter more.
READ MORE: QUANTUM SIGNALS
Quantum’s Problem Is Too Many Qubits, Not Too Few · 4 Sep 2026 · Why error-correction overhead, not qubit count, is the bottleneck.
Bitcoin’s Quantum Reckoning: Freeze It or Lose It? · 11 Jul 2026 · The exposed-key problem and the proposals to fix it.
Why The Worst Qubit Decides What A Quantum Chip Can Do · 7 Oct 2026 · A blind study of 22 qubits and the manufacturing features behind variation.
The Year Quantum Computers Learned to Fix Their Own Mistakes · 31 Jul 2026 · Error correction moves into commercial machines, still too small to win.
Disclosure: Anthropic-linked authors research and it uses an Anthropic AI model.
Sources
arXiv paper 2610.22367: authors, Table I results, method and stated limits.
Anthropic research post, 16 October 2026: the company’s account; extra context.
The Quantum Insider, 17 October: trade coverage.
Crypto Times, 17 October: cryptocurrency framing.
XenoSpectrum analysis: discussion of baseline comparisons; its disputed numbers are not used here.
The Quantum Insider on quantum threat estimates, 31 March 2026: background reading, not the basis for an attack estimate here.
The Quantum Insider on migration timelines, 8 May 2026: background reading; specific policy deadlines are not stated here.
Frontier Signal explains frontier technology in plain English. Prepared with AI assistance. Results are simulations and resource estimates from a preprint, not hardware demonstrations. This is general information, not investment advice.