A single prompt just solved a problem that stumped mathematicians for three decades. OpenAI's GPT-5.6, codenamed "CDC," produced a novel proof for a convex optimization problem involving non-self-concordant barrier functions that had remained open since the 1990s. The proof was verified in Lean 4, a formal proof assistant. This is not AlphaProof solving an Olympiad problem with specialized architecture. This is a general-purpose language model generating original mathematics at the research frontier.

The Problem That Stumped Human Mathematicians

The specific problem GPT-5.6 solved concerns the existence of a barrier function for a class of non-self-concordant barriers in convex optimization. Convex optimization is the mathematical backbone of machine learning, control systems, financial modeling, and operations research. Barrier functions are mathematical constructs that keep optimization algorithms inside feasible regions, like guardrails on a highway. Self-concordant barriers, introduced by Yurii Nesterov and Arkadii Nemirovski in the 1990s, have well-understood properties and are widely used in interior-point methods. But non-self-concordant barriers lack those guarantees. Mathematicians had been trying since the 1990s to prove that a specific class of these functions actually exists.

They failed. Not because the problem was poorly framed, but because the proof required constructing intermediate lemmas that no human researcher had thought to chain together. The problem sat in the "open problems" category for thirty years, occasionally revisited, never cracked. Until GPT-5.6 looked at it.

How GPT-5.6 Cracked It: The /Goal Paradigm

What makes this breakthrough significant is not just the result but the method. GPT-5.6 was prompted with the problem description using a new reasoning mode called the "/goal" flag. This mode instructs the model to autonomously decompose a complex goal into intermediate subgoals, and then construct lemmas that satisfy each subgoal before assembling them into a complete proof. The model generated its own intermediate mathematical statements, verified each one against Lean 4's type checker, and iterated until the full proof held.

This is fundamentally different from previous AI mathematics work. DeepMind's AlphaProof specialized in theorem proving using a custom architecture trained on formal math. GPT-5.6 is a general-purpose LLM that happens to produce formal proofs as output. It did not need a specialized training regimen for this problem. It used the same underlying architecture that generates code, writes emails, and answers customer support tickets. The /goal flag simply changed the reasoning strategy from "generate an answer" to "decompose, verify, and assemble."

The proof was run through Lean 4, an interactive theorem prover used by mathematicians to formalize proofs in a machine-verifiable format. Lean 4 confirmed the proof was correct. This verification step is critical because it eliminates the possibility of hallucination LLMs are notorious for generating plausible-sounding but incorrect mathematics. The Lean 4 verification provides formal certainty that the proof logic is sound.

Why This Matters Beyond Mathematics

If a general-purpose LLM can solve an open research problem in pure mathematics by autonomously constructing subgoals and verifying intermediate results, the boundary between "AI assistant" and "AI researcher" has effectively dissolved. The implications cascade across multiple domains.

For legal reasoning, the same chain-of-subgoal approach could decompose a complex legal argument into statutes, precedents, and logical implications each verified against a knowledge base. For financial modeling, constructing provably correct derivative pricing models becomes an exercise in goal decomposition rather than manual derivation. For chip verification, the process of proving that a hardware design satisfies its specification is mathematically identical to what GPT-5.6 just demonstrated for convex optimization.

The /goal paradigm could become a standard API primitive. Right now, most LLM interactions follow a prompt-response pattern: the user asks, the model answers. The /goal flag introduces a planning-verify-assemble loop that runs autonomously. For founders building AI applications involving formal reasoning, this paradigm shift means their products can move from "helpful suggestion" to "verified proof" without changing models, only changing prompting strategy.

What This Means for Builders

For founders building with LLMs, this result changes the strategic calculus in three specific ways.

First, formal verification integration is becoming table stakes. The proof is not credible until a formal system like Lean 4 or Coq verifies it. Any AI product claiming to produce trustworthy results in regulated domains legal, medical, financial will need a formal verification backend. Building that integration now, while it is still a differentiator, matters.

Second, the /goal decomposition pattern is portable. You do not need GPT-5.6 to implement it. The same principle autonomous subgoal decomposition with intermediate verification can be applied with any capable model using chain-of-thought prompting augmented with tool-use for verification. The competitive advantage is not the model but the architecture of the reasoning loop.

Third, mathematics as a domain has been breached. For years, the argument against AI replacing knowledge workers was that AI could not do original research, could not discover new truths, could not contribute to the sum of human knowledge. That argument is now falsified. If a general LLM can produce a verified novel proof in a thirty-year-open problem, every domain that depends on formal reasoning legal reasoning, regulatory compliance, insurance underwriting, audit, contract analysis is vulnerable to the same pattern.

Watch for OpenAI to release /goal as a public API parameter in the coming months. Watch for Lean 4 and Coq integration to become standard in AI development stacks. And watch for the first research papers listing GPT-5.6 as a co-author in top mathematics journals. That day is closer than most academics expect.