OpenAI Model Generates 722 Math Proofs

OpenAIs Unreleased Model, an internal frontier system that has not yet shipped publicly, generated 722 mathematical manuscripts from a single prompting session, according to OpenAI’s published repository in 2026. Those 722 manuscripts are not rough drafts or partial attempts. They are structured, written mathematical works organized into 372 families of related results, produced by a model that attempted roughly 4,000 problems in total, per OpenAI (2026). For practitioners tracking AI reasoning, this result sits in a different category from benchmark scores on competition math. It suggests the model is doing something closer to what a research mathematician does: constructing arguments, organizing findings, and generating knowledge that can be independently checked.

OpenAI published the collection in a public OpenAI math repository on GitHub and paired it with a community post describing the work as a first look at outputs from an internal frontier model. The scale of the output, combined with a measurable formal verification rate, makes this one of the more concrete demonstrations of AI-driven mathematical research available to the public today.

Inside OpenAIs Unreleased Model: One Prompt, 722 Proofs

OpenAIs Unreleased Model is an internal frontier AI system not yet released to the public, used to generate and organize formal mathematical results at scale from a single prompting session. That definition matters because it sets up what the 722-manuscript figure actually measures.

The core claim is striking even before you examine the content. A single prompting session produced 722 manuscripts, which OpenAI organized into 372 families of related mathematical results, according to their 2026 repository. That grouping matters because it indicates the model was not scattering disconnected outputs. It was generating thematically and structurally related work, the kind of organization a human researcher applies when writing a paper that proves several connected lemmas before arriving at a main theorem.

The model attempted roughly 4,000 mathematical problems in total, per OpenAI (2026). The 722 manuscripts that emerged represent the outputs that met some threshold of completeness or interest. The others presumably failed, stalled, or produced results the model or its evaluators judged unworthy of inclusion. That selectivity is itself a signal: the system was not just producing text. It was making something resembling editorial judgments about mathematical output.

For context, a productive human mathematician might publish a handful of papers per year, each containing one to a few proofs. Generating over 700 structured manuscripts in a single session, even granting that many are shorter or more routine than a journal submission, represents a scale of mathematical output with no clear human parallel. Whether the quality matches human research output is a separate question, and one the verification numbers begin to answer.

How One Prompt Generated 722 Mathematical Manuscripts

The session architecture behind OpenAIs Unreleased Model has not been fully disclosed. What the repository and community post reveal is that the model was given mathematical problems and produced written manuscripts as output, some of which were then checked by automated proof verification tools.

Formal verification means encoding a proof in a language that a proof assistant can check mechanically. This is different from checking whether an answer is numerically correct or whether a proof sketch looks plausible to a human reviewer. Tools like Lean check whether every logical step in a proof holds within the system’s framework, with no gaps permitted.

According to OpenAI (2026), 162 of the 722 manuscripts include a computer-checked main result, meaning those specific proofs cleared the formal verification bar. The remaining manuscripts contain results written in natural-language mathematical style, which is how most published mathematics works. Human peer review, not machine checking, is typically how those get validated. The 162 figure sets the floor for how many outputs are unambiguously correct in a machine-checkable sense.

The single-prompt framing deserves scrutiny. In practice, “a single prompt” likely means a carefully engineered instruction that initiates a long autonomous generation session, not a brief question typed into a chat interface. OpenAI has not published the full architectural specifics, but the output volume suggests a highly autonomous generation process. Any team trying to replicate or extend this work will want those details when they become available.

What the 722 Manuscript Collection Contains

The 372 families the manuscripts are grouped into span multiple areas of mathematics. OpenAI’s public repository lets researchers examine the subject matter directly. A complete subject-by-subject breakdown was not provided in the initial release, so the full disciplinary scope requires examining the repository itself.

One specific result drew separate attention: the same internal research program reportedly produced a result that disproved a conjecture in discrete geometry, according to OpenAI’s own published post on that finding. Discrete geometry deals with combinatorial properties of geometric objects, and open conjectures in the field often stand for years before resolution. A disproof, if independently verified, is a genuine research contribution, not performance on a known benchmark.

The manuscripts vary in depth and length. Some likely establish routine results; others appear to address harder problems. The 162 formally verified manuscripts are catalogued within the repository, giving researchers a clear starting point for evaluating the strongest outputs. For practitioners interested in AI-assisted mathematical research, that subset is the most immediately usable, since their correctness is machine-confirmed rather than pending human review.

The public availability of the full collection is itself significant. It allows independent researchers to assess the novelty of results, check for overlap with existing literature, and examine how the model structures mathematical argumentation. That kind of open access is what turns a press release into a research contribution.

How Researchers Verify AI-Generated Proofs

Verification is the crux of why this result matters and why the 162-manuscript figure deserves special attention. Mathematical proof has a precise standard: an argument is valid if every step follows logically from established premises. Automated theorem proving extends that standard to machine-checkable form.

When a proof is verified in a formal proof assistant, a machine has confirmed the logical chain holds completely. This is not plausibility checking or pattern matching. It is a step-by-step mechanical audit of the argument. The standard is high: every inference must be justified within the system’s logical framework.

When OpenAI reports that 162 of 722 manuscripts include a computer-checked main result, that claim means those proofs were encoded in a formal verification system and passed its checks. That’s a meaningful quality signal. It doesn’t imply the problems are the hardest in mathematics, but it does mean those solutions are correct in a verifiable, reproducible sense that any researcher can confirm by running the same verification tool against the same proof.

The broader 722-manuscript collection still requires human review to assess correctness, novelty, and significance. But having 162 formally verified results in a single collection gives the research community a rare opportunity: a large, machine-confirmed sample of AI-generated mathematical work available for analysis and use.

Mathematical Reasoning at Scale: Lessons From OpenAIs Unreleased Model

The significance of OpenAIs Unreleased Model producing this output is partly about scale and partly about category shift. AI systems have improved on mathematical benchmarks for years, but benchmark improvement is different from generating novel mathematical knowledge. Benchmarks test whether a model can answer questions with known answers. Generating a proof of something not previously proven, or disproving a standing conjecture, is a fundamentally different task.

The 722-manuscript result suggests AI systems may be approaching a threshold where they can contribute to mathematical research as generators of new results, not just as tools that help humans explore existing problems. The 162 formally verified results are the clearest evidence of this, since they represent contributions that meet the same machine-verification standard formal mathematicians use for high-confidence work.

For researchers and developers building on AI reasoning capabilities, the result also signals something about generation architecture. Producing 722 coherent, structured manuscripts organized into 372 thematic families implies a level of coherence over long outputs that most current production models don’t consistently demonstrate. That gap between what an unreleased internal model achieves and what’s publicly available is worth tracking closely.

The practical implication for mathematical research communities is that AI tools may soon need to be evaluated not just on accuracy scores but on proof-generation capacity, formal verification rates, and the novelty of outputs relative to existing literature. The OpenAI repository gives the community a concrete artifact to begin that evaluation work right now.

The 722-manuscript release gives the broader AI and mathematics communities something concrete to examine: machine-generated proofs organized into thematic families, with 162 formally verified results providing a machine-checkable quality baseline. Whether the broader collection holds up to deep mathematical scrutiny, and when this model reaches public release, remain open questions. What the data makes clear is that AI-driven formal mathematics has moved beyond isolated demonstrations. A public repository now exists for researchers to analyze, critique, and extend, which is exactly the kind of accountability structure this work needs to earn lasting credibility.

Frequently Asked Questions

How many math proofs did OpenAI’s unreleased model generate?

According to OpenAI (2026), the model published 722 mathematical manuscripts from a single session, organized into 372 families of related results. Of those, 162 manuscripts include a computer-checked main result, meaning formal machine verification. The model attempted roughly 4,000 problems in total to produce that output.

Did the model generate all 722 proofs from a single prompt?

OpenAI describes the output as coming from a single prompting session. In practice, that likely means a carefully engineered instruction that initiated a long autonomous generation process rather than a simple chat message. The full architectural details of how the session was structured have not been publicly disclosed.

Are OpenAI’s AI-generated proofs verified?

162 of the 722 manuscripts have a computer-checked main result, meaning those proofs were formally verified by a machine proof assistant, per OpenAI (2026). The remaining manuscripts are written in natural-language mathematical style and would require human peer review to assess correctness, as is standard for most published mathematics.

What subjects do the 722 mathematical manuscripts cover?

OpenAI organized the manuscripts into 372 families of related mathematical results, but a complete subject-by-subject breakdown was not provided in the initial release. The public repository allows researchers to examine the content directly. The same research effort also produced a finding that reportedly disproved an open discrete geometry conjecture, per OpenAI’s separate post on that result.

What does this achievement mean for the future of AI mathematical reasoning?

It suggests AI systems are approaching a threshold where they can contribute novel results to mathematical research, not just solve known problems. The existence of 162 machine-verified proofs in a single collection means the research community now has a large, formally confirmed sample of AI-generated mathematics to analyze, build on, and use as a quality benchmark for future systems.