New research credits and university grants support projects whose scientific results still need testing
BunCosmos Research Lab | 29 October 2026
Anthropic and Meta announced separate expansions of scientific research support on 28 October. Anthropic highlighted partnerships involving 20 foundations, institutions and research groups. Meta announced $2 million in grants across five US universities, alongside access to its models and research collaboration. Both initiatives widen access to AI; their scientific benefits remain questions for the projects to test. [1, 2]
Anthropic broadens an existing programme
Anthropic’s AI for Science programme began in May 2025. The announcement expands institutional participation, rather than starting a new programme. Partners include the Sloan and Simons foundations and the Wellcome Sanger Institute. [1]
Funders and scientific societies will distribute Claude credits to researchers, while institutions and specific projects receive support directly. Named projects include FlyWire in neuroscience, the gnomAD genetic-variation resource and Tau Ceti’s library of computer-checkable mathematics. [1]
Credits cover service use. They can make analyses more affordable, but differ from unrestricted cash or ownership of a model. Anthropic does not state their total monetary value. It says many collaborations remain early, with public datasets, software and standards among the intended outputs. [1]
Meta backs university projects with different research needs
Meta’s five partners are Caltech, the Ohio State University College of Medicine, Louisiana State University, Indiana University and UTHealth Houston. The $2 million is the combined grant amount across the five schools, not $2 million for each. [2]
The proposed medical applications are varied. An LSU project is developing computer vision to detect displaced tracheostomy tubes and signs of distress in infants. Indiana University researchers aim to make conversational systems more responsive to people living with dementia. An Ohio State project is investigating whether routine medical data can help predict a particular form of heart failure earlier. [2]
These are research objectives. Each project will need evidence appropriate to its purpose before anyone can conclude that the resulting system improves routine care. The announcements establish funding and a research direction; clinical value requires a separate test.
Some projects are designed to challenge plausible sounding answers
Caltech separately confirms support for five faculty members, including grants for up to three years and access to frontier Muse models through Meta’s API. Its projects ask questions that go beyond whether an AI can produce convincing scientific prose. [3]
Chemist Sarah Reisman plans to use unpublished laboratory notebook records to build a reaction benchmark. Such material can test predictions on evidence that was not already available in published papers. Physicist Nai-Chang Yeh is developing a framework that constrains a model’s descriptions of physical systems, so that plausible language has to connect to measurable quantities and physical principles. [3]
Anima Anandkumar’s work on TorchLean explores how agents can construct and repair scientific and machine-learning artifacts that can be formally checked. These are proposed research programmes, rather than completed demonstrations that models can independently conduct reliable science. [3]
Access is a starting point for evidence
The two announcements show companies placing their tools inside established scientific institutions, where specialists can define worthwhile questions and examine the results. Access may support useful work; it can also shape which tools researchers learn and build around. Long-term value will depend partly on what remains usable when donated access or initial funding ends.
For medical projects, the meaningful outcomes will include validation on appropriate data and evidence that a system improves a real decision. For chemistry, a proposed reaction should survive experimental testing. For mathematics, the task includes checking that a formal proof represents the intended claim.
The grants and partnerships create opportunities to perform those tests. Their success will become clearer through published methods, usable research outputs and results that other scientists can inspect and reproduce.
Sources
[1] Anthropic — AI for Science expansion 28 October 2026. Original list of 20 foundations, institutions and research groups, and the different routes for providing Claude credits.
[2] Meta — Partnerships with five US universities 28 October 2026. The $2 million grant total, model access and proposed medical research.
[3] Caltech — Five faculty selected for Meta research partnerships 28 October 2026. Recipient institution confirms model access, grants for up to three years and the specific scientific projects.