OpenAI has released 722 mathematics manuscripts, allowing outsiders to examine research produced by its model one result at a time. A complete external review is still missing, and the manuscript count does not establish how many problems have been solved. Assessing whether AI increases research output also requires knowing which statements hold, which results already existed, and how much checking researchers contribute.Release Researchers’ response
The official announcement is dated 2026-10-06 without a time or timezone. This article is dated October 7 in Taipei, with a research cutoff of 15:55, and conservatively uses the previous day’s announcement within the 48-hour window. Interesting Engineering labels its report October 6 at 18:32 EST. Using the timezone printed on that page gives October 7 at 07:32 in Taipei. That is the reporting time, not an established release time.Announcement Report
The company says the manuscripts cover 372 problem families, arising from an internal research model’s attempts at roughly 4,000 problems. On average, each result consumed computation equivalent to about three hours of ChatGPT Pro thinking with that model. Manuscripts, families and attempted problems are different units: dividing 722 by 4,000 would not establish a success rate. The model remains unreleased. What is available is inspectable research output, not a new service that ordinary users can already call.Experimental scope Compute equivalent Independent reporting
The GitHub repository provides manuscripts, some reasoning summaries and Lean formalization files, and identifies human editing of some material. For example, family 017 concerns the irrationality exponent of π, with a corresponding manuscript and reasoning summary. Outsiders can start checking a particular statement. File availability does not mean every manuscript has a complete formal proof or that outsiders have reproduced all the results.Public artifacts
Connect each proof to the question it answers
Lean can check whether a formal statement follows from its stated assumptions. Establishing whether that statement matches the original problem, and whether the result is novel, still requires subject expertise. AGMAI explicitly says its advisory role does not endorse the results’ impact or the process that produced them; assessment belongs to the mathematical community. These checks affect whether researchers can cite a result; successful compilation cannot replace them.Verification scope Community statement
For a research tool, I would present each manuscript with the original question, the statement actually proved, and the checks already completed. Suppose a model proves a claim under additional conditions. That can still be useful, but an interface that merely says “solved” could lead the next researcher to mistake a partial result for a full answer. This is a product judgment drawn from the release, not a reported user outcome.
A review record should also retain unresolved states. A missing step that needs an additional proof calls for different follow-up work from a fundamentally unsuccessful approach. Showing the specific gap gives researchers a better starting point than placing every unconfirmed manuscript on a list of successes.
Count review time alongside additional research output
OpenAI’s release increases the material outsiders can inspect directly. AGMAI’s response also makes clear that expert review remains work to be done. For teams considering AI-assisted research, I would favor tools that preserve evidence and revisions over a choice based only on how many manuscripts they generate at once.Availability Review status
Suppose two systems produce the same number of usable results. If one preserves assumptions and failed attempts clearly enough for colleagues to establish more quickly which steps can be reused, it may be better suited to collaborative research. That choice depends on checking actually taking less time, rather than passing unfinished verification to someone else.
The public information does not yet establish a total cost per novel result. The three-hour figure describes a model-compute equivalent, not a measurement of full human review. What can accumulate next is a record of individually confirmed, novel results and the work required to confirm them. That would help establish whether publishing more research artifacts lets mathematics reach reusable findings sooner.Compute and limits External review
The cover reuses the image from this site’s earlier OpenAI mathematics article. It is not a new image from this release; downloading an event image failed because DNS resolution was unavailable.
Sources: