A statistics professor asked GPT-5.6 Sol, the slow-thinking version of OpenAI’s newest AI, to examine a math question that had stayed unsolved for thirty years, and the AI settled it in about 90 minutes. It is the clearest sign yet that the AI in your pocket now comes in a version built for genuinely hard problems.
The Gist
- A Wharton professor used GPT-5.6 Sol Pro to knock down a statistics rule trusted since 1995
- The previous AI, GPT-5.5, tried the same problem for about 20 hours and failed
- The professor published the proof, the code and the full chat so anyone can check it
- The regular GPT-5.6 everyone can use is not this exact version, but the family is the same
Have ChatGPT Recap This Article
ChatGPTMeet Sol, the Slow-Thinking Version of GPT-5.6
GPT-5.6 is the latest AI model from OpenAI, the company behind ChatGPT, and it ships in several versions with different speeds. GPT-5.6 Sol is the one that takes its time to think, and the Pro variant used here gets even more thinking room for each question.
Thinking room has a concrete meaning for an AI. Before answering, the model writes out reasoning steps for itself, a bit like scratch paper, and the Pro setup simply allows far more of that scratch paper per problem. You already met this family when GPT-5.6 rolled out worldwide earlier this month.
That extra patience is exactly what a hard math question needs. Quick answers work for emails and recipes. Open problems need thousands of careful steps that all have to hold together.
How One Professor and 90 Minutes Rewrote a Math Rulebook
Edgar Dobriban, a statistics professor at the Wharton School in Pennsylvania, gave Sol Pro the formal definition of a statistical method and one question about it. Roughly an hour and a half later, the AI had built a working answer that experts had chased for three decades.
The scale of the upgrade is easy to feel. The previous model, GPT-5.5, ground away at the same question for about 20 hours and got nowhere. GPT-5.6 Sol needed 90 minutes, and a hands-on reviewer already put Sol and its faster siblings Terra and Luna through a full test when they launched.
One detail makes this story trustworthy rather than hype. The professor published the result, the code and the entire chat with the AI, so any mathematician can replay the conversation and verify every step.
Keep learning on AI Noobies:
- Spotify Music AI Now Talks to You in the App
- Every K-12 Teacher in the US Just Got Claude for Free
- The New Siri Is Now Open to Every iPhone in the Beta
What a Conjecture and a False Discovery Rate Actually Are
A conjecture is a math statement everyone believes is true but nobody has proven. This one dated back to 1995 and concerned a method called Benjamini-Hochberg, a tool scientists use to avoid fooling themselves when they test thousands of ideas at once.
That method promises to keep the share of false alarms, called the false discovery rate, under a chosen ceiling. The AI found a situation where the ceiling gets broken: a target of 0.1 that actually lets 0.104 through. A tiny crack, but in a tool cited more than 130,000 times, which in science is celebrity status, small cracks matter.
Zooming out, the interesting part is how the AI did it. It did not invent new math. It combined known techniques in a way nobody had tried, the same kind of stubborn recombination that showed up when students’ grades collapsed once AI help was removed: these tools are now deeply woven into academic work, for better and for worse.
How to Watch This Kind of AI Reasoning Yourself
You do not need the Pro version to see slow thinking in action. In ChatGPT, pick the thinking model when it is offered and ask something with real structure, like planning a three-generation family trip with a fixed budget. Watch how the answer builds in ordered steps instead of one quick blurb.
The difference with the free tier is mostly patience. The heavy version gets more time and more scratch paper, which is why it is sold to researchers and companies first, and why access to GPT-5.6 is being handed out in stages.
The honest takeaway for a beginner: AI just moved from summarizing knowledge to occasionally producing new knowledge, in public, with receipts. That is worth remembering the next time someone tells you these tools only repeat what they read.
Stay tuned on AI Noobies.
About the Author
