Anthropic’s Claude Cuts a Quantum Decoder’s Gates by 12x

Anthropic said on 16 October 2026 that Claude Mythos 5, given plain-language specifications for eight quantum circuits and an automated evaluator, found designs that improve on the best published ones: “Working largely on its own, Claude found designs that improve on the best published circuits, in some cases by an order of magnitude or more.”

In one case, Anthropic says the changes “reduced the number of gates in the decoder by a factor of 12 relative to Google’s previous construction”. In another, it says one point-addition design “uses 12% more qubits but 17x fewer gates” than the smallest previous design, and that design came from a single 53-hour run.

Business Pill · AI EVALUATION

A one-minute explainer of AI evaluation as an acceptance test: how to know an output is good enough. It teaches the general idea only and says nothing about any company in this story.

What Anthropic Reported

The work was led by Andi Gu as part of Anthropic’s STEM Fellows Program, with mentorship from Anthropic quantum physicist Alek Kemeny, in collaboration with Susanne Yelin of Harvard University and Mikhail Lukin of Harvard University and QuEra Computing, according to Anthropic’s post. The paper, arXiv 2610.22367, was submitted on 15 October 2026.

The eight tasks span chemistry, optimization, quantum simulation, number theory and numerical computing, Anthropic says. The paper’s abstract states: “In seven of the eight tasks, the discovered circuits significantly reduce both Toffoli and qubit counts relative to the state of the art.” In the eighth, it says, the search found a new reversible architecture for prime-field modular inversion.

The score Claude worked to improve was “the number of qubits used, multiplied by the number of gates, which serves as a proxy for how long the circuit would take to run”, the post says. Toffoli gates are among the most expensive gates to run, so counting them is a standard way to estimate run time.

What Was Posted on X

Anthropic announced the research on X on 16 October 2026: given plain-language instructions, it says, Claude Mythos 5 designed quantum circuits that improve on the best published ones, sometimes by an order of magnitude or more, and it links the research post this piece draws on.

News account Watcher.Guru, on 16 October, relayed Anthropic’s announcement that Claude Mythos designed highly efficient quantum circuits that could bring quantum computing ‘within reach sooner’, quoting Anthropic’s phrase.

How the Loop Worked

According to Anthropic, Claude wrote each circuit as Python code and sent it to an evaluator program, which checked whether the design was correct and scored it on qubits and gates. The score went back to Claude, which used it for its next attempt, with no human input between rounds.

For point addition, the core step in quantum algorithms for the discrete logarithm problem on elliptic curves, Anthropic says Claude got a one-paragraph brief. “After working for 53 hours with minimal human input (limited to the initial brief and occasional prompts to think more broadly), Claude arrived at a completely new approach to the problem.”

Anthropic says Claude proved several new algebraic identities and checked its decoding formula with large-scale numerical simulations and in Lean, a proof language. “With the full validation complete, Claude produced the first reversible implementation of divsteps with no stored branch history.”

The Decoder Result

The second result Anthropic describes is decoded quantum interferometry, or DQI, an optimization algorithm whose most expensive step is a Reed–Solomon decoder. Anthropic says estimates by Google Quantum AI researchers imply that one such problem would take about three months on the Frontier supercomputer, against roughly an hour on a future quantum computer running DQI.

Claude switched the decoder to a method called Berlekamp–Massey and reconstructed most temporary values instead of storing them, the post says. “Setting aside enough memory for the worst possible case, in which an exceptional value arises in 35 different rounds, would take 665 qubits.”

Instead, the post says, “Claude wrote a high-speed simulator and ran 403 million sample decodings” to measure how often exceptions occur. “So Claude’s circuit sets aside room for just one saved value, or 19 qubits.” Anthropic adds: “About one run in 10,000 needs more than that and fails, which is well within the error rate the task allows.”

Where Anthropic Saw Limits

Anthropic lists limitations. “In our early experiments, Claude was conservative by nature.” It says Claude second-guessed apparent improvements and tended to prefer well-known approaches, so the setup required several different approaches in parallel and the team occasionally prompted it to broaden its search.

Anthropic compares the pace with past progress on FeMoco, a chemistry benchmark: “the 150,000-fold drop in the FeMoco estimate described above took many research groups eight years”, while “By contrast, Claude’s elliptic-curve design was the result of a single 53-hour run.” Earlier this week we wrote about Google’s Willow work on quantum magic-state errors, and about Bank of America hiring contractors to fix bugs Anthropic AI found.

The Structural Read

As we read it, the task was chosen because it can be checked: Anthropic says each input to these circuits gives one definite output that an ordinary computer can verify, so an external evaluator could validate every proposal.

As we read it, the gains came from changing the method, not tuning: Anthropic says Claude replaced the extended Euclidean algorithm with Berlekamp–Massey decoding and reconstructed temporary values instead of storing them, cutting the reserve to 19 qubits.

As we read it, the human role moved to the brief and the nudges: Anthropic says the 53-hour point-addition run had minimal human input, limited to the initial brief and occasional prompts to think more broadly.

Three Implications

QUANTUM HARDWARE AND ALGORITHM TEAMS Anthropic says fewer qubits and gates reduce the quantum resources useful applications need; the paper says seven of eight discovered circuits cut both Toffoli and qubit counts versus the state of the art.

AI LABS AND RESEARCH GROUPS The setup Anthropic describes pairs a model with an automated checker and a single score, qubits times gates, with no human input between rounds and a Lean proof for the divsteps design.

ANYONE READING THE CLAIMS Anthropic reports limits: Claude was conservative early on, second-guessed improvements and needed prompts to broaden its search; one decoder design accepts about one failed run in 10,000.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework argues: “Consolidation wins where work is digital, modular, and verifiable”.

As we read it, quantum circuit design as Anthropic describes it fits that profile: the circuits are code, an evaluator checks them, and a single score ranks them. The framework predates the announcement and does not discuss quantum computing.

What Is Not Established

The post and paper report results against previously published designs; we did not find an independent replication in the texts we read.

We read Anthropic’s research post and the arXiv paper on 16 October 2026. We did not contact Anthropic, Harvard or QuEra, and we did not check the circuits ourselves.

The Bottom Line

Anthropic said on 16 October 2026 that Claude Mythos 5, working in a loop with an automated evaluator, designed quantum circuits that improve on the best published ones across eight tasks, cutting the gates in one decoder by a factor of 12 and producing a point-addition design with 17x fewer gates for 12% more qubits.

The paper says seven of the eight discovered circuits reduce both Toffoli and qubit counts. Anthropic also reports limits: Claude was conservative early on and needed prompts to pursue more novel approaches.