Unreleased OpenAI model takes on 10 major math problems — first

We don't have the weights, the name, or even a release window, but a new OpenAI model reportedly sat down and cracked 10 major unsolved mathematical problems. That's not a benchmark with a few extra percent on GSM8K. That's territory where a model stops pattern-matching and starts doing something that looks a lot like research.

The list of problems isn't public in full detail, but the claim itself is what matters. Math is the cleanest test of reasoning we have — no ambiguity, no hidden internet retrieval. If a model can produce genuine proofs for open problems, it means the frontier is no longer about memorized knowledge. It means the thing actually reasons.

Let me be clear about what this does and doesn't tell us.

• Verification gap: "Solved" is a strong word. For an open problem, the proof has to be checked by mathematicians. A model can generate a plausible-looking argument that falls apart at step 40. The real signal is whether these proofs withstand peer review.
• Where the ceiling is: This is likely a post-training effort — massive reinforcement learning on formal math systems, then search at inference time. That's the same recipe behind the recent o3-style results: spend more compute at test time, let the model explore proof branches.
• Not just token prediction: If this holds, it's evidence that scaling test-time compute beats scaling pretraining for hard logic. Compute favors the side with a verifier. Math has one built in.

People will ask whether this is a genuine advance or just a stunt. To me, the smart play is to look at which problem areas were solved. If they're algebra or combinatorics, that's one thing. If they touch analysis or number theory where proofs are long and nonlinear, that's another. Long-horizon reasoning is precisely where all the current models choke.

The deeper point: we've spent a year talking about agents, MCP, tool use and Claude Code-style workflows. This result, if real, drags the conversation back to the core question — how good is the reasoning engine itself? Because an agent that can browse your repo is cute, but an agent that can prove a theorem is a different species.

I also think this is a strategic signal from OpenAI. Releasing news of an unreleased model solves nothing for users today, but it resets expectations. The narrative changes from "incremental gains over last year" to "the next step is a qualitative jump." It also puts pressure on labs like DeepMind and xAI to show their own unannounced reasoning work.

What I'd watch for next is not the model name but the evaluation setup. If the problems come with formal verification over the whole proof, we're in a new regime. If it's just natural language output with high confidence, be skeptical. The difference is the difference between arithmetic and real math.

Either way, the bar for "AI reasoning benchmark" just moved. Someone needs to release a public benchmark with genuinely open problems, one where the ground truth isn't known — because that's the only way to test what this feels like.