OpenAI Releases 722 Math Papers: Pricing Impact

OpenAI published a large collection of AI-generated mathematical work on October 6, 2026. Its new mathematics repository lists 722 manuscripts across 372 families of related results. The company says an unreleased internal frontier model produced the work and that many—but not all—proofs have accompanying Lean formalizations. This is a research disclosure, not a model launch or a price change.

What OpenAI released

The repository contains an overview, manuscripts with citation and revision instructions, a formalization catalogue, and 10 abridged summaries of the model’s reasoning. OpenAI says it posed approximately 4,000 problems during the evaluation. It reports that the average result used compute equivalent to roughly three hours of ChatGPT Pro thinking with the internal model. The figure is an estimate of work, not an invoice: the model is not publicly available and OpenAI did not publish its operating cost, token ledger, or customer price.

Item	What is public	What remains uncertain
Manuscripts	722 papers in 372 families, with citation and revision instructions	Whether every claim survives mathematical scrutiny
Formal proofs	Lean artifacts for many results	Not every manuscript is formalized
Process	10 abridged reasoning summaries and approximate problem/compute counts	Full reasoning traces, per-result token usage, and dollar costs
Model	Described as an internal frontier model	Public access, release date, API name, and rate

OpenAI’s own README cautions that some unformalized results could have issues and says corrections will be recorded as new versions. A Lean artifact can make a particular formal statement machine-checkable; it does not automatically validate every surrounding interpretation, originality claim, or unformalized manuscript. The repository should therefore be read as material for expert review, not a blanket claim that 722 open problems are solved.

Pricing impact: no customer rate change

The live table above shows existing public API list prices for GPT-6 Luna, GPT-6.1 Sol, and GPT-6 Astra. It is a comparison baseline, not a price for the unreleased research system and not evidence that one of those public endpoints can reproduce the collection. OpenAI’s announcement does not state an old or new API rate, a ChatGPT Pro plan change, or when the internal model might be offered to customers.

Buying question	Before this release	After this release
Public OpenAI API rates	Existing rate card	No announced change
Research model access	Not publicly available	Still not publicly available
Cost per verified research result	No comparable public bill	Still undisclosed

Do not multiply the three-hour estimate by a ChatGPT subscription fee to derive a paper cost. A Pro subscription is not a metered hourly research-model tariff; retries, failed attempts, verification, human editing, and formalization would also affect a real project’s economics. For a separately built research assistant, start with our OpenAI pricing page and token calculator. Compare another public provider on the Anthropic pricing page, and see our GPT-6 model selection guide for what the currently available tiers actually do.

What this means for researchers and buyers

Mathematicians gain access to inspectable manuscripts, revision history, and some machine-checkable proofs. They also inherit a substantial review burden: the repository spans hundreds of papers and explicitly says some may need corrections. Research groups should prioritize claims relevant to their field, check the exact theorem and assumptions, and record which repository revision they reviewed.

AI buyers gain a capability signal, not a procurement option. The model’s availability, rate limits, contract terms, and price are unknown. Teams building scientific agents should evaluate currently accessible models on their own tasks and count verified results, failed attempts, expert review time, and compute spend—not manuscript volume alone.

For a separately hosted open-model control, compare Novita’s current model catalog using the same problems and expert-review standard. Its models are not the private OpenAI research system.

Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored link at no extra cost to you.

There is also a governance dispute. OpenAI says it consulted the independent Advisory Group on Mathematics and Artificial Intelligence and is exploring a community-hosted repository. The group’s September 29 recommendations explicitly oppose testing advanced math on proprietary models and call for independent scholarly hosting, process disclosure, formalization where feasible, and support for human understanding. Publishing on OpenAI’s GitHub and promising workshops address parts of that agenda; they do not settle the access or independent-hosting concerns.

What to do now

1. If a result matters to your work, read its manuscript and formalization status in the repository catalogue; do not rely on the headline count.
2. Pin the cited revision and seek domain-expert review before using a claim in a paper, product, or security-sensitive decision.
3. If you are budgeting an AI research workflow, benchmark a publicly callable model and track cost per expert-accepted result. Do not budget an unreleased model from the Pro-equivalent estimate.
4. Watch for an actual model-access announcement and official rate card before treating this research as a new buying option.

Bottom line: OpenAI has made an unusually large body of AI-generated math inspectable. The model and its economics remain private, and mathematical validation will take time. Today’s news changes the research record—not the public API price list.

Sources: OpenAI’s October 6 announcement, the OpenAI mathematics repository, the official API pricing documentation, and the advisory group’s September 29 recommendations. Checked October 7, 2026.