Skip to content

Anthropic

Claude Mythos 5 beat the best published quantum circuits

6 min read

Our AI analyst checked this story itself: it ran, measured and compared against the originals. What we checked

Written by an AI model from the linked sources and checked against them by a second model. How we work

Image: anthropic.com

Claude Mythos 5 designed quantum circuits that beat published designs by 4.1x to 39x on six benchmark tasks, scored as qubits multiplied by Toffoli gates, according to a paper Anthropic released on October 16. The other two tasks don’t produce a clean multiple, and the most-cited figure from the release, 17x on elliptic-curve point addition, takes some context to read.

The work, described in Anthropic’s research post and an accompanying arXiv paper, was led by Andi Gu in Anthropic’s STEM Fellows Program. Alek Kemeny, a quantum physicist at Anthropic, mentored the project, and Susanne Yelin (Harvard University) and Mikhail Lukin (Harvard University and QuEra Computing) collaborated on it. The eight tasks come from chemistry, optimization, physics simulation, number theory and numerical computing. The paper says seven of the resulting circuits use fewer Toffolis and fewer qubits than their baselines, and the eighth offers a new trade-off between the two.

How the loop works

Claude got a plain-language specification of each circuit and an automated evaluator. It writes a design in Python. The evaluator checks that the output is correct and scores it on qubits times gates, a stand-in for runtime, and that score feeds the next attempt with no human input between rounds. The paper says the search ran on OpenEvolve, with several families of designs developed in parallel and a fresh agent session for each iteration.

There is a catch: the loop only works for algorithms where each input gives one definite output that an ordinary computer can check. Many useful quantum routines fall into that class, but not all of them.

Gate counts matter because they set how long a computation runs. The Toffoli gate, a three-qubit operation that is among the most expensive to run, is the standard unit for that estimate. In 2017, computing the ground-state energy of FeMoco, the iron-sulfur cluster at the core of the nitrogen-fixing enzyme nitrogenase, was estimated to need about 50 trillion Toffolis. Later work brought that down to about 340 million, a roughly 150,000-fold drop. Most of that came from redesigning circuits rather than from better hardware. We checked these figures against the 2025 tabulation by Low et al.: 341 million Toffolis, with 3.2 million physical qubits running for 2.7 hours.

Quantum circuits make that redesign hard in two ways. A circuit can’t branch the way an ordinary program does. It runs on a superposition of inputs, so it has to include the operations for both sides of every “if”. It also can’t throw away scratch work. Intermediate results have to be uncomputed, which means running the steps that produced them in reverse, and that costs gates. Keeping them around instead costs qubits.

Where the gains came from

Anthropic says each improvement came from a new structural or mathematical insight about the problem, rather than from fine-tuning an existing design. In an iron-sulfide molecule simulation, Claude used the molecule’s symmetries and repeating structure to compute some results once and reuse them. In other circuits it removed costly components that only handled extremely rare inputs. The paper’s tables allow a failure probability of at most 10<sup></sup>-3 per task for this, and 2·10<sup></sup>-4 for modular inversion.

The clearest single case is decoded quantum interferometry (DQI), which Anthropic calls a leading quantum optimization algorithm. Claude replaced the extended Euclidean algorithm in its decoder with Berlekamp-Massey decoding. That cut the decoder’s gate count 12x against Google Quantum AI’s earlier construction and made the algorithm’s most expensive step a little over an order of magnitude cheaper. The Quantum Insider reported the counts as about 440,000 Toffolis, down from 5.1 million, with peak qubits falling from 1,938 to 1,837. The outlet also notes that these are circuit designs and resource estimates, not calculations run on a working quantum computer.

The point-addition run

Point addition combines two points on an elliptic curve to produce a third. It is the core step in quantum algorithms for the elliptic-curve discrete logarithm problem, and a standard benchmark in the field. Claude got a one-paragraph brief that told it to explore several approaches in parallel. Anthropic says it then worked for 53 hours, with human input limited to the brief and occasional prompts to think more broadly.

Most of the cost sits in modular inversion, the wraparound version of division: given a, find x such that ax ≡ 1 (mod p). Classical computers do this efficiently with divsteps, which repeatedly swaps, adds, subtracts or halves a pair of numbers. A quantum version has to keep a record of which operation it chose at every round so the computation can be reversed, and storing that record takes a very large number of qubits. Claude found a way to avoid storing the full history, which gave a new reversible circuit for the inversion. Anthropic says one circuit in that family uses 12% more qubits but 17x fewer gates than the smallest previous design.

Our analysis

What we checked

On October 18 we compared the arXiv paper with Anthropic’s post, recalculated each task’s gain from the paper’s figures, and ran the authors’ public evaluator on three circuits.

How we check stories

Anthropic’s post says all eight designs beat the best published circuits on both qubits and gates. The paper says seven, and one of those seven is compared against the authors’ own naive implementation. The 17x point-addition gain holds, but only against the published design with the fewest qubits. Against the one with the fewest Toffolis, Claude’s circuit loses on the combined score.

TaskGain, qubits × Toffolis (paper)
Fixed-point Gaussian39x
HHL rotation angle32x
Reed-Solomon DQI decoder12.2x (11.6x on Toffolis alone)
Coulomb potential9.9x
Lattice gauge7.7x
Fe2S2 alias-Select4.1x
  • Point addition: 4,006,673 Toffolis and 935 qubits, against 69,529,757 and 838 for Luo et al. We also compared it with Schrottenloher’s circuit, using the figures the paper reports (that circuit isn’t in the code release). Claude’s design needs 2.1x more Toffolis and 35% fewer qubits, and its combined score is 1.36x worse.
  • The paper describes the Hermitian DQI baseline as a “Naïve reversible implementation; no prior reversible compilation exists in the literature.”
  • The 53-hour figure appears only in Anthropic’s post.
  • Our evaluator runs on the Coulomb-potential and HHL circuits and on the Coulomb baseline matched the paper’s counts exactly, with zero wrong outputs. That reproduced the 9.9x Coulomb gain end to end.

Running the evaluator

The circuits, task definitions and exact evaluator are published under Apache-2.0 at andigu/uncompute. The baselines are reconstructions of published constructions. Each of our three runs took 3 to 6 seconds on one CPU with the pure-Python evaluator. The README warns that large circuits take minutes to hours on a single machine. A small task runs with uv run python tools/eval.py coulomb-potential.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.