OpenAI published new results on open problems in mathematics from an internal frontier model on October 6, 2026, releasing 722 manuscripts organized into 372 families in a public GitHub repository alongside Lean proof formalizations. OpenAI, Unite.AI, and Interesting Engineering covered the release. The core model remains unreleased, so outside mathematicians cannot yet reproduce the results independently. The proofs shipped. The model did not.
I run ten production platforms. The construction ERP at coenconstruction.com validates material cost escalation clauses in change orders. The estimating tool at estimate.pro reconciles subcontractor bid submissions against historical pricing distributions. The Punchlist at thepunchlist.vercel.app flags inspection items that map to local permit requirements. Every workflow calls a model. Every model integration started with the same question: can I test this model on my data before I route production traffic to it. OpenAI pushed 722 mathematics manuscripts to a public GitHub repository on October 6, 2026, crediting the work to an internal frontier model the company has not released or named. I can clone the repository. I cannot call the model.
The repository contains artifacts but no endpoint
According to the README, the vast majority of results were obtained with the same procedure using an unreleased internal OpenAI model, and over the course of the evaluation, the model was posed approximately 4,000 problems, with the average result using the equivalent compute of roughly three hours of ChatGPT Pro thinking. The dog-training site at valhalla-k9.com generates session summaries from trainer notes. The freelancer portfolio at hiremariaelena.com scores intake-form answers against project-fit criteria. The permit-sketching tool parses zoning ordinances into constraint sets. Each integration runs in a loop: load sample inputs, call the model, measure the outputs, compare the results to the previous baseline, decide whether to ship. As part of the GitHub repository, OpenAI is sharing formalizations of many of the proofs in Lean, a programming language that allows mathematical proofs to be checked by a computer, and will update the repository with more formalizations as they obtain them. I can read the Lean files. I cannot run inference against the model that produced them.
The construction ERP ships schedules of values to lenders. The estimating SaaS produces bid packages that general contractors submit to procurement offices. Curbscript generates service routes that field crews drive. The platform can call itself enterprise but the API does not always support the workflows that enterprise operators run. The workflow I need this week is offline comparison. On October 6, OpenAI shared results from tests on about 4,000 math problems, covering areas like number theory and algebraic geometry, with many including Lean formalizations for verification. The formalization lets me verify the proof. It does not let me test the model on a new problem.
Lean checks logic but does not grant access to the reasoning system
A critical component of this release is the use of the Lean programming language, a tool designed for computer-verified mathematics that allows computers to check the logical consistency of a proof. The three Base44 internal apps process payroll classifications, route service tickets, and reconcile vendor invoices. Each app started as a manual workflow. Each manual workflow turned into a model call when I could measure the model output against ground truth and prove the error rate stayed below the threshold the process could tolerate. Lean can confirm that a chain of logical steps follows the rules once someone has translated an informal proof into its strict syntax, but it cannot confirm that the informal statement being formalized is interesting, novel, or correctly describes the problem the paper claims to solve. Lean tells me the proof is valid. It does not tell me whether the model can solve the next problem I give it.
The model can train on the tools I use but the API does not always expose the capability the training unlocked. The capability I need is reproducibility. Lean formalizations cover the main result of 162 of the 722 papers, per the repository's formalization catalogue. One hundred sixty-two proofs are machine-checkable. Seven hundred twenty-two manuscripts came from a model I cannot access. I can verify twenty-two percent of the work the model did. I cannot ask the model to do new work.
The model wrote the proof. I can check the proof. I cannot call the model.
The construction ERP parses change-order PDFs. The estimating tool extracts line-item descriptions from Excel attachments. The permit sketcher converts hand-drawn markup into constraint polygons. Every parser started as a test: load a sample file, call the model, diff the output against the expected result, measure the delta. OpenAI says it will be funding a series of workshops, conferences, and special programs around the understanding of major results produced by AI, and that it is working to responsibly release the model that produced these results, describing the model only as an internal frontier model. OpenAI will release the model later. The manuscripts are public now. I can clone the proofs today. I cannot test the reasoning system that generated them until OpenAI decides the model is ready to ship. The repository went live October 6. The model did not.