PSI Used OpenAI and Claude Models on 5 Open Quantum Problems

Physical Superintelligence (PSI), a physics research company that says it has raised $58M+, announced on X on 26 October 2026 “solutions to five long-standing problems in quantum information theory.” Each of the five results comes with a paper, a Lean formalization and verification code, its research page says.

PSI’s release repository lists the AI models used on each result, and every one names models from both OpenAI and Anthropic, including OpenAI’s GPT-6-Astra on all five. The announcement came the same day OpenAI said its own system had refuted the Exponential Time Hypothesis.

Video explainer

A Business Engineer explainer on PSI’s results and the week’s AI-assisted proofs, narrated from our published pieces.

What PSI Announced

The results address five entries in the Open Quantum Problems list, numbers 2, 12, 27, 38 and 43. PSI says its researchers obtained them “working with our AI research system”, and that the release repository “documents how AI was used in each.” The research page is dated 23 October.

On OQP 12, PSI says it proves that no choice of local observables violates the CHSH inequalities beyond a finite separation between two regions of the vacuum, “even though the two regions remain entangled at arbitrary separation.” On OQP 43, it says it constructs a bipartite counterexample to the claim that every such causal process can arise from a larger finite-dimensional reversible process.

Not every result closes its problem. For OQP 2, PSI shows Bell pairs cannot be distilled from three copies of certain Werner states, and writes: “The question for arbitrarily many copies remains open.”

What Was Posted on X

Physical Superintelligence PBC posted on 26 October 2026 at 23:44 UTC that it is “announcing solutions to five long-standing problems in quantum information theory”, linking its research page. The page itself calls them “new results”.

Andrew Curran quoted PSI’s post at 01:02 UTC on 27 October, writing that PSI “launched in September with a $58 million seed” and listing the five problems: Open Quantum Problems 2, 12, 27, 38 and 43, each with a paper and a Lean formalization.

Which AI Models PSI Used

Each result guide has a section titled “Provenance and use of AI”. Across the five, PSI lists OpenAI’s GPT-6-Astra, GPT-5.6, GPT-5.6 Pro and GPT-5.6-Sol, and Anthropic’s Claude Fable 5.1, Claude Opus 4.8, Claude Opus 5 and Claude Opus 5.5 for the mathematics.

The Lean formalizations, the guides say, were written by AI agents under PSI’s direction: GPT-6-Astra in Codex on all five, and Claude Opus 5.5 in Claude Code on four. Named PSI researchers, among them David Turturean, Ning Bao and Ian MacCormack, are listed as responsible for the correctness of each paper.

One result builds directly on an OpenAI finding. For OQP 38, PSI notes that two PPT maps do not suffice in general, “as reported in a recent OpenAI preprint”, and says it proves that 2(d − 1) such maps do.

Who PSI Is

PSI describes itself as a public benefit corporation building “the world’s most advanced physics research lab”. Its site lists $58M+ in funding led by Breakthrough Energy Ventures, and 25+ researchers and engineers. Andrew Curran wrote on X that the company launched in September with a $58 million seed.

Its site says “Human scientists direct the missions. Virtual physicists carry out the research,” and names the physical infrastructure of AI as its first proving ground.

Connecting the Coverage

Overnight we reported that Anthropic says Claude also refuted ETH, on 16 October, the same date OpenAI gives for its claim; on 26 October we reported that OpenAI says its AI refuted the Exponential Time Hypothesis. In September we covered how Claude formalized Fermat’s Last Theorem in Lean.

The Structural Read

As we read it, the disclosure format is the notable design choice: each result guide names the models, says which agents wrote the Lean formalization, and names the PSI researchers responsible for the paper’s correctness.

As we read it, the release lands in the same week OpenAI and Anthropic each said their own internal models refuted the Exponential Time Hypothesis, so AI-assisted proofs are arriving from the labs and from a lab customer at once.

As we read it, OQP 38 shows results stacking: PSI’s bound builds on a finding it credits to a recent OpenAI preprint.

Three Implications

FOR PHYSICS RESEARCHERS Each result folder contains a paper and a Lean 4 formalization, and each lean/ folder has its own pinned dependencies and can be copied out as a standalone project, PSI’s README says.

FOR MODEL VENDORS PSI’s result guides list OpenAI’s GPT-6-Astra on all five results and Anthropic’s Claude models on all five, by our count of the guides.

FOR INVESTORS PSI’s site lists $58M+ in funding led by Breakthrough Energy Ventures, and names the physical infrastructure of AI as its first proving ground.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The framework describes “The Verifier Moat” as one “built by RLVR on domain-specific reward signals.”

As we read it, a Lean formalization is that kind of verifier for mathematics, and PSI ships one with every result. The framework was published in July 2026 and does not discuss PSI.

What Is Not Established

PSI’s repository describes the five papers as preprints. The Lean files check the formal statements written for them, which the guides map to specific manuscript versions. PSI’s X post calls them “solutions”, while its research page calls them “new results”, and the OQP 2 result covers three copies only. The PSI documents we read do not cite an outside review.

The Bottom Line

PSI says AI-assisted workflows produced results on five open problems in quantum information theory, each with a paper and a Lean formalization. Its repository lists models from both OpenAI and Anthropic on every result, with named PSI researchers responsible for each paper.