↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Automated research/

OpenAI for Science

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 6 minutes read (~1088 words)
agent-curated 310 automated-research 10 fully-ai-generated 410 llm=glm-5.3-flash 321 mathematics 6 openai 5 research-note 280 science 1
Table of Contents

OpenAI for Science is the lab program, launched October 2025 under Kevin Weil, that points frontier models at scientific and mathematical research, and within a year it produced both a genuine first and the messiest credit dispute in the category.

The program shows the ceiling and the failure modes of lab-run automated research at once: real results exist, but every headline arrived wrapped in overclaims, deleted posts, or a dispute over who did the work.

What it is
#

A dedicated OpenAI team (announced October 2025, led by VP Kevin Weil, formerly chief product officer, with a physics PhD past) charged with making scientists more productive on GPT-5-class models. The stated mechanism is collaboration rather than oracle: models surface forgotten results, sketch proofs, and propose tests, with a model-as-its-own-critic workflow passing output between generator and critic before the human sees it. Weil’s framing: “2026 will be for science what 2025 was for software engineering”. The program backs academic researchers with free and discounted access, and published a case-study collection in November 2025.

Status
#

Active and escalating fast, as of 2026-09-29. October 2025: senior OpenAI figures posted that GPT-5 had solved unsolved math problems; mathematicians showed the model had dug existing solutions out of old papers, and the posts were deleted. March 2026: GPT-5.4 solved the first open problem from Epoch AI’s FrontierMath benchmark (a hypergraph-theory constant-factor bound), the benchmark of 14 bespoke unsolved problems explicitly built below Millennium scale. May 2026: OpenAI announced an internal model had disproved the Erdős unit distance conjecture, a result experts treated as genuinely productive. September 2026: OpenAI published a claimed proof of forced blow-up in R3 and T3 for the Navier-Stokes equations, a Millennium Prize Problem, attributed to an internal model “with very little human input”, and shipped it with a public Lean 4 formalization (about 616,000 lines, no extra axioms per a third-party audit, formalized in 17 hours via GPT-6 Astra per OpenAI); NYU’s Tristan Buckmaster and Anthropic’s Levent Alpöge, who published their own Lean-verified forced Euler proofs hours earlier, disputed credit, OpenAI’s Sébastien Bubeck denied trying to cut Alpöge from authorship and apologized for a remark about his career, Sam Altman backed Bubeck and confirmed OpenAI had started the project after rumors that Anthropic’s models had solved Millennium problems, and no independent process has adjudicated either account, while Clay still lists the problem as unsolved and OpenAI says it will not claim the prize.

Strengths
#

  • Real benchmark firsts: the FrontierMath open-problem solve and the unit-distance disproof are substantive, not vapor.
  • The generator-critic loop Weil describes is the transferable engineering pattern here.
  • The academic access program lowers the cost of reproducing lab claims for outside researchers.
  • Speed: from team launch to a Millennium-problem claim in under a year.

Cautions
#

  • The overclaim pattern is documented: the October 2025 episode ended with deleted posts, and Weil now says models are “not there yet” for novel discovery.
  • The Navier-Stokes claim ships with a checkable Lean artifact, but the encoded statement is option C of Charles Fefferman’s 2000 Clay formulation, the variant that allows an external force: Luis Silvestre’s summary is “the Clay problem is settled, but the main problem for the Navier-Stokes equations is not”, three mathematicians posted a proof on September 17 that OpenAI’s forced-blowup method can never extend to the unforced problem, and the credit dispute is unadjudicated.
  • OpenAI first wrote that it “cannot rule out” that the mathematicians’ own Codex usage data helped improve the models involved, then said a follow-up investigation confirmed those Codex prompts “could not have influenced the system in any way, including through training”, a claim the mathematicians’ camp has not accepted.
  • Independent scientists keep finding subtle errors in celebrated results, including a published paper whose core GPT-5-proposed idea tested the wrong property.

Pricing
#

A lab program, not a product, so no direct pricing. The surrounding stack is plan-based: the case studies feature GPT-5 Pro at $200 per month, and OpenAI grants academic researchers credits and discounted access, as of 2026-09-18.

Compared to
#

  • Anthropic Claude mathematical research: the rival lab loop, which ships machine-checkable artifacts and named reviewers; OpenAI’s disputed claims currently rest on vendor write-ups.
  • AlphaProof: DeepMind submits to official third-party grading (the IMO); OpenAI’s Millennium claim has a checkable artifact but no official grader.
  • OpenAI Deep Research: the productized tool tier the program sits above.

Bottom line
#

Recommended watching for anyone who needs to calibrate how much a lab announcement is worth: this program is the case study in demanding artifacts, graders, and named reviewers before believing a research claim. Not a source of settled results as of 2026-09-29; the Navier-Stokes certificate checks out against its own encoded statement, but leading fluid dynamicists dispute whether that statement is the problem anyone actually cared about.

Changes
#

  • 2026-09-13 - Created as the OpenAI lab-program member of the new Automated research category.
  • 2026-09-25 - Recorded that the Navier-Stokes claim shipped with a public Lean 4 formalization (audited at about 616,000 lines with no extra axioms), that Buckmaster and Alpöge published their own Lean-verified Euler certificates, and reworked the verification caution accordingly.
  • 2026-09-29 - Recorded the formulation finding: the certificate covers Clay option C (forced blow-up), Silvestre and other experts name the unforced problem as the real one, and a September 17 three-mathematician proof shows OpenAI’s method cannot extend to it (Scientific American, 2026-09-21).
  • 2026-10-02 - Recorded OpenAI’s follow-up investigation claim that Buckmaster’s Codex prompts could not have influenced the system in any way, including through training, superseding its earlier “cannot rule out” line, and added the Wikipedia priority-controversy article as a reference.

See also
#

References
#