Simons, OpenAI and Anthropic Put $15 Million Behind Mathematics. Read the Verb: It Funds Checking the Work, Not Doing It.

On October 15, 2026, the Simons Foundation announced that it, Anthropic and OpenAI will commit a combined $15 million in funding and resources to help the mathematics community deal with artificial intelligence. Each of the two AI companies is putting in $5 million. The Simons Foundation will manage and award the money, including through an open call for proposals.

Nine days earlier, on October 6, OpenAI had released more than 700 manuscripts at once — 722 of them, organized into 372 groups of related results — claiming solutions to hundreds of open mathematical problems. The reaction from mathematicians ranged from disorientation to anger. Fields Medalist Hugo Duminil-Copin wrote, "I am paralysed." Terence Tao acknowledged that the proofs appeared to contain genuinely clever new ideas and said he was nonetheless deeply frustrated by the absence of anyone who could discuss the work, present it in a seminar, or teach it to students. The Association for Human Mathematics issued a statement saying that "Mathematicians did not ask for this work to be done" and accusing OpenAI of a "total disregard for the norms of scientific research."

Understanding that sequence is the key to reading the grant program, because the program is not shaped like a research fund. It is shaped like an absorption mechanism.

What was announced

Total: $15 million in combined funding and resources.

Split: $5 million from OpenAI, $5 million from Anthropic, with the Simons Foundation contributing the balance and administering the whole thing.

Who awards it: The Simons Foundation, exclusively. The companies are funders, not grantmakers, in this structure.

What it funds: Workshops, conferences and special programs, plus an open call for proposals supporting work that helps researchers evaluate recent AI-assisted results, learn and use new tools, share perspectives, and shape collective standards for AI use.

In-kind, separately: Anthropic and OpenAI will also provide AI tools and computing resources to Simons-supported researchers — and that component explicitly extends beyond mathematics into basic science.

Timing: Announced October 15, 2026. Detailed submission guidelines for the open call were not published with the announcement.

Foundation President David Spergel framed the purpose broadly: "Now is the time to look at mathematics as a whole — not as a collection of proofs and theorems, but as a collective understanding driven by human curiosity."

The fundable unit is a convening, not a theorem

Grant seekers reading the headline will reasonably assume this is money to do mathematics with AI. It is not, and the distinction determines whether your application is responsive or dead on arrival.

Every activity named in the announcement is infrastructure for collective sense-making: workshops, conferences, special programs, standards. The open call sits alongside them, and the described purposes — evaluate results, learn tools, share perspectives, shape standards — are all second-order activities performed on mathematical output rather than activities that produce it. The Simons Foundation already runs substantial first-order programs for doing mathematics: Simons Investigators, Simons Collaborations, Targeted Grants in the Mathematics and Physical Sciences, and the Simons Symposia series. This new money is not a competitor to those. It is a new category.

That makes the natural applicant pool wider and stranger than the usual Simons audience:

• Formalization and proof-assistant groups. The Foundation was an early backer of the Lean proof assistant, which it credits with changing how mathematicians work and with making AI-assisted proofs checkable at all. A program about evaluating AI-produced results flows naturally toward the one technology that mechanically verifies them. If you run a formalization effort, a mathlib contribution pipeline, or a tool that translates informal proofs into machine-checkable form, you are near the center of this program's intent.
• Scholarly societies and journals. The 722-manuscript episode created an immediate, unresolved editorial problem: what does peer review mean for a submission whose author cannot answer questions about it, and what disclosure does a journal require. Standards work needs a convener with standing. That is a society, not a lab.
• Departments and institutes that want to run a workshop. Modest, concrete, high-fit. A three-day workshop on verification practice for a subfield is exactly the shape of thing this money was raised to pay for.
• Research-integrity and priority-dispute work. Two mathematicians disputed OpenAI's September claim about the Navier-Stokes equations, suggesting the company had built on their unpublished work — a charge OpenAI denies. Whatever the merits in that case, attribution for AI-assisted results is now a live, unsolved methodological question, and methodology questions are fundable.
• Training and curriculum. Tao's complaint was pedagogical as much as anything: there is no one to teach the result. Programs that build teaching capacity around machine-generated mathematics address a stated need from the most-cited critic.

If your draft concept note's main verb is "prove," "classify," or "construct," you are applying to the wrong Simons program. If it is "convene," "verify," "formalize," "evaluate," "standardize," or "train," you are in the right one.

Who is paying, and why the structure matters more than the amount

Two-thirds of this money comes from the companies whose output the mathematics community is being asked to adjudicate. That is worth saying plainly, because it will be the first objection raised in any department that debates whether to apply, and because the program's design anticipates it.

The mitigation is in the governance: the Simons Foundation manages and awards the funds. OpenAI and Anthropic do not select grantees. That is the standard structural answer to a corporate-funded independence problem, and it is a real one — it is the same reason an endowed chair funded by a company is treated differently than a consulting contract with one. It is not a complete answer. A funder that writes a $5 million check into a $15 million pool shapes what gets proposed even without touching the review panel, simply by making a category of work visible and resourced.

For an applicant, the practical consequence is twofold. First, write the disclosure into the proposal rather than around it. A standards project funded partly by the vendors whose systems the standards govern is defensible if and only if the governance is explicit: who sits on the body, how dissent is recorded, what happens when a funder disagrees with an output. Reviewers at a foundation that just absorbed this reputational question will read governance sections carefully.

Second, note what the companies are contributing besides cash. AI tools and compute, provided to Simons-supported researchers across basic science, is an unpriced component sitting outside the $15 million headline. In-kind compute is genuinely valuable and genuinely awkward. It usually arrives as credits on a specific platform with its own terms of service, data-retention policy and usage caps, and it typically cannot be valued, cost-shared, or carried over the way cash can. If your institution requires in-kind contributions to be documented at fair market value for its own reporting, start that conversation before you accept credits, not after.

What this says about where AI-and-science money is going

Set this next to the rest of October 2026 and a pattern emerges. The Department of Energy launched a Genesis Mission Graduate Fellowship pilot and named twelve Phase II awards worth $159 million. Biohub, DOE, NIH and three AI labs committed $1.8 billion to a virtual biology initiative. The Rockefeller and OpenAI foundations opened a $50 million program for AI in city government. Nearly all of that is money to build AI capability.

The Simons announcement is the first large commitment in this cycle pointed in the opposite direction — money to metabolize AI capability that already exists and arrived faster than the receiving field could handle. That is a distinct funding category, and if the 722-manuscript episode repeats in physics, chemistry or biology, it is a category that will grow. Organizations that build credible competence in evaluating machine-generated scientific output are positioning for more than one open call.

There is also a narrower read worth holding: $5 million is a rounding error relative to what either company spends on training runs, and the donation followed the backlash rather than preceding it. Observers who describe this as reputational expenditure are not wrong. Reputational expenditure still funds real workshops, and a grant seeker's job is not to adjudicate the funder's motives. It is to notice that a new pot exists, understand its shape, and be ready when the call opens.

What to do now

The submission guidelines are not out. That creates about a month of useful preparation time before anyone can actually apply.

Draft a one-page concept now, in the program's own vocabulary. Name the community you convene, the specific standards or evaluation capacity that does not currently exist, and what artifact exists at the end — a published protocol, a verified corpus, a trained cohort, a disclosure norm adopted by a journal. Vague community-building proposals will lose to specific ones, because the Foundation is buying a solution to a problem it can name precisely.

Identify your co-signers early. Standards work is only worth funding if the people who would have to follow the standard are in the room. A proposal with three society officers and two journal editors attached is a different document than the same idea from one department.

Watch simonsfoundation.org for the call, and assume a relatively short window once it posts. Foundations responding to a live controversy do not usually run twelve-month lead times.

And decide your institution's position on the in-kind compute before you need it. The cash is uncomplicated. The credits are not.