OpenAI

The directions that can currently be confirmed are: a domain model for the life sciences (GPT-Rosalind) and its biodefense branch, a science benchmark OpenAI published along with the gap it has not closed (FrontierScience), a scientific writing surface (Prism), a disclosure of how far agents have entered its own research organization, Codex driving real laboratory instruments, and a mathematics claim that arrived inside an authorship dispute. Unlike DeepMind, there is almost no "build a family of science models" route here: apart from GPT-Rosalind, what appears in the public record is general models plus tools, interfaces and access.

Scope note: this page tracks the 2026 public record only: official releases, papers, benchmarks, third-party reporting and named collaborator disclosures. No inside information. It records how far public material goes, not how well the organization is doing.

Overview

1. Life-science model: GPT-Rosalind

What can be confirmed: on April 17, 2026 OpenAI released GPT-Rosalind, described as its first domain-specific model series, aimed at biochemistry, genomics and protein engineering, supporting "evidence synthesis, hypothesis generation, experimental planning, and multi-step scientific workflows." It is named after Rosalind Franklin.

The numbers: a 0.751 pass rate on the BixBench bioinformatics benchmark; better than GPT-5.4 on six of eleven LABBench2 tasks. A third-party test run by Dyno Therapeutics placed it above the 95th percentile of human experts on sequence prediction and around the 84th percentile on sequence generation.

How it is released: a trusted-access programme, U.S. enterprise customers first, where applicants must show they are working toward improving human health outcomes. A Life Sciences research plugin for Codex connects the model to more than 50 scientific tools and data sources.

Boundary: the gated release is itself a response to dual-use risk: reporting notes researchers' warnings that models trained on biological data could help design dangerous pathogens. Confidence: high (official release plus a named third-party test). But BixBench and LABBench2 are computational benchmarks, not wet-lab output, and Dyno's test covers sequence tasks, not drug outcomes.

Open leads: how many organizations trusted access actually reached; what each of the four launch partners (Amgen, Moderna, Thermo Fisher Scientific, the Allen Institute) did with it; whether any wet-lab replication is public.

2. Biodefense: Rosalind Biodefense

What can be confirmed: on May 29, 2026 OpenAI announced Rosalind Biodefense, giving vetted developers access to GPT-Rosalind to build pandemic-preparedness and biosecurity tools, plus a government track for public-health workflows. Named partners include Lawrence Livermore National Laboratory, the Johns Hopkins Applied Physics Laboratory and the Coalition for Epidemic Preparedness Innovations.

Source wording: "Frontier AI should meaningfully advantage those defenders."

Boundary: this is an access programme, not a capability result. The public record contains no output, evaluation or deployment result for any defensive tool built on it. Confidence: medium (official announcement, no results disclosed).

3. Science benchmark: FrontierScience

What can be confirmed: released December 16, 2025, with a paper in January 2026. Two tracks: Olympiad (international-olympiad difficulty, 100 questions in the open gold set) and Research (PhD-level open-ended research sub-problems, 60 questions, split evenly across physics, chemistry and biology). Olympiad problems were written by 42 former international medalists or national-team coaches holding 108 olympiad medals between them; research problems by 45 post-doctoral researchers, professors or doctoral candidates.

Results: GPT-5.2 scores 77% on Olympiad and 25% on Research; Gemini 3 Pro scores 76% on Olympiad. The 52-point spread is the most informative thing in the benchmark: the same models are near saturation where the answer is well defined and far from it where the problem is open.

Scoring: olympiad answers are matched against a single numeric or algebraic expression; research answers are graded against a 10-point rubric by a model judge, with seven points counted as success.

The paper's own four limitations (verbatim):

• "A significant part of scientific research is proposing novel research directions, hypotheses, and ideas", which the benchmark does not evaluate;
• "The rubric is less objective than the equivalency checker of a single expression or number, and relies on the model judge's capabilities";
• problems are text-only, "without image or video outputs";
• "We did not perform human baselining of this dataset and leave that to future work."

Confidence: high (paper, open gold set, stated limitations). An organization publishing the gap it has not closed is more checkable than any single score it reports.

4. Scientific writing: Prism

What can be confirmed: released January 28, 2026. A LaTeX workspace combining drafting, citations and model assistance, free to ChatGPT users.

Boundary: the public record is a product release: no usage, retention or research-output data. Confidence: medium (official release, no usage data).

Why it is tracked: this is a move into the daily research surface. Writing and citation are where evidence gets organized, cited and passed on; holding that surface is a different thing from model capability.

5. Agents inside the research organization

What can be confirmed: on September 6, 2026 OpenAI reported that agents can complete well-defined tasks that would take skilled researchers several days, and disclosed that by mid-August its research organization was running 3.1 agent-workdays for every human workday.

Boundary: this is company-reported internal usage with no external audit and no statement of what those workdays produced. Usage is not output. Confidence: medium (self-reported, single source).

Open leads: how the ratio is computed, whether it counts failed runs and retries, and whether any externally checkable definition exists.

6. Instrument control: Codex on a superconducting chip

What can be confirmed: on September 8, 2026, an MIT researcher connected the model through Codex to laboratory software, letting agents choose parameters, operate qubit measurements, analyze results and set the next measurement.

Boundary: noisy or ambiguous signals still require expert judgment. This shows agents entering a live measurement loop; it does not show unsupervised experimental autonomy. Confidence: medium-high (official technical post, named institutional collaboration).

7. Mathematics: the Navier–Stokes claim and the authorship dispute

What can be confirmed: on September 8, 2026 OpenAI said an internal model produced an analytical proof and Lean formalization for the Millennium Prize formulation, constructing a finite-time singularity under smooth forcing. The same day, Tristan Buckmaster (NYU) and Levent Alpöge (Anthropic) said their earlier unpublished forced-Euler work had already reached OpenAI, and that OpenAI resisted listing Alpöge because of his Anthropic role; Sébastien Bubeck disputed parts of that account, said the methods differ, and apologised for his wording.

Boundary: what is recorded here is a claim and a dispute, not a settled mathematical result. The accounts differ and the proofs remain under review. Confidence: low (single-source claim, public dispute, review incomplete).

Why it gets its own section: at the frontier of mathematics a capability claim now arrives bundled with questions of priority, data access and authorship, and peer review was not designed to settle those inside a news cycle.

8. National labs and policy

What can be confirmed: since January 2025 OpenAI has worked with the U.S. National Laboratories, with up to 15,000 scientists able to use its reasoning model series, and with Microsoft deployed a model on Venado, the Los Alamos supercomputer. OpenAI is among the AI partners named in the Department of Energy's Genesis Mission, and filed a response to the White House OSTP request for information on accelerating science.

Organizational signal: "OpenAI for Science" is led by former chief product officer Kevin Weil, now VP of Science. The stated hiring plan is a "small team" of academics who are both world-class in their fields and "completely AI-pilled," and good science communicators, paired with OpenAI researchers.

Confidence: high (official releases plus government-side public material).

What could not be found

• How many organizations trusted access actually reached, and what the four launch partners produced with it;
• The definition behind the 3.1 ratio, in terms anyone outside could check;
• Usage or research output for Prism;
• Any independent review verdict on the Navier–Stokes proof;
• The size, membership and project list of the OpenAI for Science team.

Method and scope

Scope: only work where AI is applied to scientific research. General model capability and consumer products are excluded.

Confidence is assigned claim by claim, not to a whole direction. Within one direction, "the project exists" and "the technology works" can have very different evidence strength. The labels:

• High: a paper, an open gold set, a third-party test, or a named institution executing the work;
• Medium: an official release or company-reported figure, inspectable in detail, unverified outside;
• Low: a single-source claim, a spoken statement, or a claim under public dispute with review incomplete.

Two operating rules: revised papers are dated by version and are not treated as new projects; evidence that a project exists is not evidence that the technology has succeeded.

Primary sources

Official releases

• Introducing GPT-Rosalind for life sciences research, 2026-04-17
• Rosalind Biodefense, 2026-05-29 (via R&D World)
• Evaluating AI's ability to perform scientific research tasks (FrontierScience), 2025-12-16
• Introducing Prism, 2026-01-28
• Research acceleration: a view inside OpenAI, 2026-09-06
• Codex and quantum computing experiments, 2026-09-08
• OpenAI: Navier–Stokes solution, 2026-09-08
• Strengthening America's AI leadership with the U.S. National Laboratories

Papers and benchmarks

• FrontierScience: Evaluating AI's ability to perform expert-level scientific tasks (arXiv: gold-set sizes, who wrote the problems, stated limitations)

Third-party reporting

• OpenAI launches GPT-Rosalind, The Next Web (BixBench, LABBench2, Dyno Therapeutics figures)
• Mathematicians dispute credit for OpenAI's Navier–Stokes result, The Next Web
• OpenAI partners with U.S. National Laboratories, CNBC
• Energy Department unveils 24 AI research partners in Genesis Mission push, FedScoop