OpenAI Releases 372 Math Results From an Unreleased AI Model

On Tuesday, Oct. 6, 2026, OpenAI published a large set of new mathematical results that it says were produced by an internal AI model that has not been released to the public. The company grouped the work into 372 families of results, drawn from 722 manuscripts, and posted them on GitHub, according to Scientific American and other reports.

What happened

According to Scientific American, an OpenAI spokesperson said the model produced almost all of the results in response to a single prompt given to a single AI agent, although some may have taken more than one attempt. OfficeChai and AI Weekly reported that each result used about three hours of ChatGPT Pro-level thinking time on average.

The results span many areas of mathematics. Reports say they include claimed progress on several famous open problems. OpenAI has not named the model that did the work.

The company was careful about what the release means. OfficeChai reported that the results have not been peer reviewed. OpenAI said many, but not all, of the proofs have been written in Lean, a programming language that lets a computer check every step of a proof. It acknowledged that some results without a computer-checked proof could contain problems, and it said it will keep adding more formal proofs.

The release drew strong reactions from mathematicians. Some were struck by the speed of progress. Others questioned how outsiders can check the work when the model itself and the exact prompts are not public.

Why it matters

Math is one of the hardest tests of reasoning. A proof is either right or wrong, and there is no room for a confident-sounding guess. If an AI model can produce real, correct new mathematics at this scale, it says a lot about how capable the next generation of AI systems will be.

That is also why the questions matter. Large claims need careful checking, and computer-checked proofs in Lean are one of the best tools for that. The results that have been formally checked are much more trustworthy than the ones that have not. Over the coming months, the math community will decide how much of this work holds up.

The release also gives a hint of where AI is going. A model that can work alone on hard problems for hours, and keep track of long chains of reasoning, could do the same kind of deep work in science, engineering and software. That is a big shift from chatbots that answer quick questions.

What this means for founders and builders

For founders, the lesson goes beyond math. AI is getting better at long, careful, multi-step work. That opens the door to products that handle real research, analysis and problem solving, not just short replies.

The story also shows why trust and checking matter so much. Any AI product that makes important claims, whether about money, health, law or learning, needs a way to show its work and let people verify it. Products that build in that trust will win over products that simply sound confident. Simple features such as showing sources, explaining each step and letting users double-check results can make a big difference in whether customers rely on an AI tool.

At KIING DIGITAL LABS, we build complete AI-powered products, including learning tools. Math Master Lab, a fully built mathematics learning platform, is available on our Acquisitions page. If you have your own idea for an AI tool, a custom build turns it into a finished product.

The bottom line

OpenAI's release of 372 math results from an unreleased model is a striking sign of how fast AI reasoning is improving. But many of the claims still need to be checked, and mathematicians want more transparency before they accept them. For builders, the takeaway is clear: AI can now tackle deeper problems, and trust and verification are what turn that power into products people rely on.

Free tool: Have an idea sparked by this news? Turn it into a product plan in minutes with Concept Studio — free. Skip the build: Explore fully built products ready to acquire or start a custom build.

• OPENAI
• AI RESEARCH
• MATHEMATICS
• AI MODELS