AI · Oct 21, 2026
OpenAI on Google Cloud lets companies spend Google commitments on GPT-6, and that changes procurementOpenAI posts 722 AI-generated math manuscripts to GitHub, partly following its advisers’ guidance
· Updated · Signal Desk newsroom
Analysis of reporting by The Verge. How we work

The Verge reports OpenAI released 722 manuscripts in 372 result families from an unreleased model, with reasoning summaries, compute estimates and attempt statistics. AGMAI says it includes solutions to hundreds of open questions, but OpenAI did not name the model and chose GitHub over academic channels.
Key takeaways
- The release has 722 manuscripts in 372 result families, per The Verge; other reports cite 719.
- AGMAI says it includes solutions to hundreds of open questions.
- The average result used about three hours of ChatGPT Pro thinking, OpenAI says.
- OpenAI published on GitHub with revision and citation protocols.
- The model is described only as unreleased.
The release
The Verge reports that OpenAI has released a batch of 722 manuscripts, grouped into 372 result families, produced by an unreleased frontier model. TechCrunch and The Verge’s later coverage cited 719 manuscripts (advisers’ standards, scale of the drop). The independent Advisory Group on Mathematics and Artificial Intelligence (AGMAI) says the release includes solutions to hundreds of open questions; in September, OpenAI had said the model resolved more than 100 long-standing open problems.
The release includes some summaries of the model’s reasoning, compute estimates and statistics on how many problems were attempted. OpenAI says the average result used the equivalent of about three hours of ChatGPT Pro thinking. It is publishing on GitHub with protocols for revisions and citations and says it is still exploring community-hosted alternatives. The release adds to a fast-growing body of AI-produced math from OpenAI and rivals, including Anthropic, with results touching a Millennium Prize problem.
Grading against AGMAI
AGMAI’s late-September recommendations asked labs to release promptly, use established academic channels where possible, disclose the model’s name, prompts and compute, and stop treating math results as marketing (AGMAI). OpenAI moved on some of that: compute estimates, reasoning summaries and attempt statistics are genuine disclosures. But the model is described only as unreleased, GitHub is not an established academic channel and OpenAI frames other venues as something still under exploration, which amounts to partial compliance.
A workload, not just a gift
If even a fraction of the results hold, this is a significant moment for AI in science. But mathematics advances as people read, check and argue over proofs, work done by a limited, mostly unpaid pool of experts on their own schedule. Hundreds of papers at once is a workload as much as a gift.
The harms if the pattern scales are concrete. Mathematicians whose work these systems build on may struggle to get credit (attribution concerns); early-career researchers could see open problems closed overnight with papers landing in a repository rather than a journal; and verification capacity becomes the bottleneck. Releasing results is better than hoarding them, but pace and format matter, and labs generating hundreds of papers could help fund the checking. Our overview tracks the broader debate (math drama). More in our OpenAI coverage and research coverage.
Who feels it
- Mathematicians
- Hundreds of manuscripts to assess, with credit and overlap questions.
- Journals
- A GitHub-first release tests traditional publishing.
- AI labs
- AGMAI guidelines are the yardstick; partial compliance gets noticed.
What to watch
- How many results survive expert review.
- Whether OpenAI names the model and discloses prompts.
- Moves to place results in journals or community venues.
- AGMAI’s public response.
FAQ
How many manuscripts were released?
722 in 372 result families, per The Verge; other reports cite 719.
Where were they published?
On GitHub, with revision and citation protocols.
What did OpenAI disclose?
Reasoning summaries, compute estimates and attempt statistics.
What did it not disclose?
The model’s name and prompts.
Companies: OpenAI
More on OpenAI
AI · Oct 14, 2026
White Circle says newer OpenAI and Google image models are easier to jailbreak than their predecessorsAI · Oct 19, 2026
Microsoft ships Guardian Agents, saying frontier models ‘can’t be trusted’AI · Oct 13, 2026
After months off the frontier, Google ships Gemini 4 Argon with benchmark claims and aggressive pricingAI · Sep 22, 2026
Opus 5.5 hands risky cyber prompts to an older Claude model, and that routing is the real launch story