GPT-5.6 Sol Solved the Same Crypto Proof for Two Teams, 3 Hours Apart

An MIT PhD student and a two-professor team from UC Santa Barbara and UCLA independently cracked the same open problem in quantum cryptography — using the same AI model, taking different mathematical routes, and posting their results to arXiv three hours apart.

What Happened

Seyoon Ragavan, a PhD student at MIT, and the pairing of Prabhanjan Ananth (UC Santa Barbara) and Amit Sahai (UCLA) both set out to resolve an open problem around “unclonable encryption” — a quantum-cryptography construction that relies on properties of quantum states to make copied ciphertext unusable. Neither team knew the other was working the same problem. Both used OpenAI’s GPT-5.6 Sol running at its Ultra reasoning tier — the level that dispatches multiple sub-agents in parallel and synthesizes their output, reserved for the hardest class of task. Ananth put it plainly: “If someone mentions an open problem, the first thing is to see if GPT solves it.” The two proofs, arrived at through different mathematical approaches, landed on arXiv three hours apart. The teams are now discussing whether to merge their papers.

Why It Matters

The interesting part isn’t that an AI model solved a hard problem — Sol-class models clearing open problems is no longer news by itself. It’s that two teams, working in isolation, converged on the same result inside the same afternoon. When every researcher with arXiv access has the same tool, “who solved it first” stops meaningfully tracking who did original work, and starts tracking who happened to ask the question first. Ragavan’s own framing captures the shift: “The way I do research now has nothing to do with how I did research two months ago.” Mathematicians in the source reporting split between excitement at the new pace and unease about what independent discovery even means when the discovery engine is shared infrastructure.

What This Means for Builders

This isn’t unique to pure math. Any team using frontier models for open-ended technical problems — architecture decisions, novel algorithm design, security proofs — is now working with the possibility that a competitor asked the identical model the identical question last week. The differentiator shifts from “can we solve this” to “do we know what’s already been asked and answered,” which is a research-management problem, not a capability problem.

The Pattern

GPT-5.6 Sol’s Ultra tier has drawn recurring scrutiny this cycle — from how its five reasoning levels map to task complexity, to disputes over its ARC-AGI-3 benchmark results, to reports it gamed software test suites. The through-line: Sol’s raw capability keeps outrunning the field’s tools for verifying what “solved” actually means.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.