Anthropic opens AI for Science to 20 foundations and institutes: Claude becomes the default infrastructure for labs

Anthropic opens AI for Science to 20 foundations and institutes: Claude becomes the default infrastructure for labs

🔎 Public science just changed providers

On October 28, 2026, Anthropic quietly executed one of the most significant strategic pivots in its history. Its AI for Science program, launched in May 2025 to fund individual researchers one by one, is now shifting to the institutional scale: 20 foundations, research institutes, and scientific consortia receive free Claude credits, with a stated goal — making Claude the default AI infrastructure for academic research worldwide.

The partner list includes heavyweights: the Simons Foundation, the Alfred P. Sloan Foundation, Schmidt Sciences, the NSF — the US federal agency that accounts for roughly 25% of federal support for basic research —, the Wellcome Sanger Institute, EMBL, the Francis Crick Institute, and Imperial College London. This is no longer philanthropic sponsorship. This is distribution.

And the timing is no accident: the announcement lands in the middle of Anthropic's push week — drug discovery, the GSA contract, the SpaceX partnership — just days before its IPO opening in November, following the S-1 filing under the ticker ANT. We analyze the announcements, the numbers, and above all the question that no one at Anthropic is highlighting: what happens to scientific independence when it runs on a proprietary model?

The Essentials

• 20 institutional partners join AI for Science on October 28, 2026: funders (Simons, Sloan, Schmidt Sciences, NSF, American Chemical Society), institutes (Sanger, Crick, EMBL, Imperial College, Gladstone Institutes, Space Telescope Science Institute), and projects (FlyWire, gnomAD, AstroPath, LEAP).
• Three distribution channels: funders redistribute credits to their grantees, institutes consume them internally, and research projects receive them directly.
• Stated goal: exceed the 10,000 seats initially planned for scientists, with premium seats at $15 per month offering 5× the standard usage limits.
• Strategic context: the third announcement of a busy week, on the eve of the November 2026 IPO — public science becomes an acquisition and lock-in channel.
• Point of friction: part of global research now depends on a proprietary model controlled by a company in an aggressive monetization phase.

The 20 partners and the three distribution streams — Direct answer

Anthropic no longer distributes credits on a case-by-case basis: it is placing Claude at the heart of the funding and execution chain of public research, through three distinct channels.

According to Anthropic's official announcement, the program is now organized into three streams. First stream: the funders. The Alfred P. Sloan Foundation, the Simons Foundation, Schmidt Sciences, or the Open Source for Science Fund (backed by Renaissance Philanthropy) receive credits that they redistribute to their grantees. Concretely, a researcher funded by the Sloan Foundation can now be allocated a Claude budget on top of their grant.

Second stream: the institutions. The Wellcome Sanger Institute, EMBL, the Francis Crick Institute, Imperial College London (Department of Life Sciences), or the Gladstone Institutes use the credits internally for their own programs. The Sanger Institute states in black and white that it intends to automate and accelerate large-scale computational biology and genomics analysis with Anthropic's platforms.

Third stream: the projects. FlyWire (Princeton), LEAP (Columbia), AstroPath (Johns Hopkins), gnomAD (Broad Institute), the Centre for Population Genomics, or Tau Ceti receive the credits directly, as identified scientific consortia.

To this is added an atypical institutional partner: the NSF, a federal agency with a $9 billion budget, alongside the American Chemical Society (CAS). When the agency that funds a quarter of American basic research sits at the table of a private model provider, the signal goes far beyond the scope of sponsorship.

What labs are actually doing with Claude — A direct answer

Behind the press release, there are concrete scientific experiments — and some are surprisingly ambitious.

At Imperial College London, several groups in the Department of Life Sciences are receiving token access. Two projects illustrate the level of ambition. The first: predicting how fungal proteins disable their host's defenses — with roughly 90% accuracy — then using AI to design proteins capable of blocking those attacks. This is a far cry from "Claude helps me proofread my thesis": we're talking about language-model-assisted protein design.

The second project is even more telling: researchers are testing whether Claude can independently rediscover a genetic transfer mechanism that was never published, using only the evidence available before the original discovery. This is a test of autonomous scientific discovery, not assistance. If the model pulls it off, the marketing pitch will write itself.

This shift toward structured research workflows had been set up by the launch of Claude Science, Anthropic's AI workbench for researchers, which turns the chatbot into a scientific working environment. The free credits are the acquisition layer; the workbench is the product; the institutional partners are the distribution channel. The loop is complete.

A Week of Strategic Push on the Eve of the IPO — Direct Response

This announcement doesn't stand alone: it's the third pillar of a coordinated week that positions Anthropic as an infrastructure provider — and it all plays out before its IPO opens in November.

Let's revisit the timeline. Anthropic announced drug discovery initiatives, signed an agreement with the GSA (the U.S. federal government's procurement agency), then unveiled this AI for Science expansion. The throughline is crystal clear: demonstrate to investors that Claude is not a niche chatbot but cross-cutting infrastructure — government, healthcare, science, enterprise.

The brand's earlier moves tell the same story. Salesforce adopted Claude as its default reasoning engine, transforming the world's #1 CRM into an agentic platform. On the compute side, Anthropic committed with SpaceX to Colossus 1: 220,000 GPUs and 300 MW dedicated to Claude — the infrastructure capacity that makes a promise of scientific scale credible. And on the open source front, Anthropic opened up SRT, Claude Code's sandbox whose restrictions are gradually becoming a reference standard.

All of this fits into the financial calendar: Anthropic filed its S-1 under the ticker ANT on the NASDAQ, with an IPO expected in November. An S-1 prospectus must tell a story of sustainable growth. "Hundreds of scientific projects supported since May 2025," as TokenPost reminds us, and now 20 institutional partners — that's exactly the kind of traction-plus-legitimacy narrative that investors pay a premium for.

My take: the philanthropy is real, but it's no longer the product. The product is the de facto standard.

The business model: from free credits to lock-in — Direct answer

Free credits are an acquisition investment, not a gift: every dollar given away today builds a billable dependency tomorrow.

The mechanism is classic and remarkably effective. Step 1: give credits to researchers whose publications cite Claude. Step 2: watch scientific workflows — analysis pipelines, extraction agents, internal tools — get built around the Anthropic API. Step 3: once the dependencies are in place, start charging. The move from free standard seats to premium seats at $15 per month with 5× the limits, detailed in the "Expanding our support for scientists" announcement, is the first rung of this pricing ladder.

Let's compare the models being distributed to measure what the labs are getting. According to the June 2025 rankings, Claude Opus 4.7 (Adaptive) scores 90 in general evaluation and 94.3 in agentic use, while Claude Sonnet 4.6 offers a good compromise at 83 for general use — enough to cover both scientific writing and multi-step pipelines. Facing GPT-5.5 (91 general, 98.2 agentic) or Gemini 3.1 Pro (92 general), Claude is not alone at the top, but Anthropic's strategy isn't played out on the benchmark: it's played out on institutional distribution. OpenAI and Google have academic programs; neither has signed the NSF, Sloan, and Sanger within the same week.

Worth noting, the expansion of the thematic scope: AI for Science is no longer limited to the biological sciences. Anthropic explicitly mentions "very compute-hungry" research, which brings the program closer to compute-heavy use cases — and therefore to higher-value customers, consistent with the SpaceX partnership mentioned above.

The critical issue: public science's dependence on a proprietary model — Direct response

This is the question the announcement sidesteps, and it deserves to be asked head-on: what becomes of reproducibility and scientific independence when part of the world's research runs on a closed model, controlled by a publicly traded company?

Three distinct problems. Reproducibility first. The scientific method rests on the ability to reproduce an experiment. But a proprietary model evolves continuously: the version of Claude used in January will no longer be available in December, responses are not deterministic, and the training context is not publicly auditable. A discovery obtained "with Claude" is not reproducible in the strict sense of the term — it is reproducible "with Claude, such version, such date." Scientific journals are only just beginning to define disclosure standards for this case.

Vendor dependence next. When an institute like Sanger builds its large-scale genomics pipelines on Anthropic's platforms, it accumulates a colossal rewrite debt in the event of price changes, changes to terms of use, or sector-specific restrictions. Businesses learned this risk with the cloud; public science, for its part, has neither a migration budget nor a vendor lock-in culture. And the temptation to change the terms increases mechanically with the IPO: a publicly traded company answers to its shareholders, not to the public's bioinformaticians.

Legitimacy capture finally. The NSF funds basic research with public money. If NSF grantees use credits from a single vendor, the US government is de facto delegating part of its scientific infrastructure to a private player. No competitor will have the same access. We're no longer in competition over models, we're in competition over channels — and Anthropic has just locked up the best ones.

Honest counter-argument: public funding of research has never been neutral. Donations from private foundations, commercial software licenses, and contracts with scientific publishers already create dependencies. And alternatives exist — DeepSeek V4 Pro (88) and Kimi K2.6 (84) offer open-weight models, GLM-5.1 (83) can be self-hosted. But none of these players has Anthropic's institutional distribution, and that's precisely what counts.

What this changes for individual researchers — Direct answer

The lone researcher remains affected, but their status changes: they go from being a direct beneficiary of the program to being a link in the redistribution chain.

Since May 2025, AI for Science targeted individual researchers — hundreds of projects supported in biology, medicine, physics, chemistry, climate, and mathematics. That channel remains open, but growth now happens through institutions: a PhD student at the Crick or a post-doc at Imperial will get their credits through their lab, not through an individual application.

For scientists not covered by a partner, the Team plan for scientists becomes the standard entry point: limited free seats, $15 per month for intensive usage. It's an aggressive introductory price — comparable to a Spotify subscription — for a market whose real value will be decided by institutional contracts and the API.

My practical advice: if your lab is among the 20 partners, use the credits for prototyping, but document your dependencies from day one. An analysis pipeline written against a single vendor's API is a liability, not an asset.

❌ Common Mistakes

Mistake 1: Believing that "free credits" means "neutral engagement"

Many commentators present AI for Science as pure philanthropy. This overlooks the fact that every scientific publication citing Claude is a marketing asset, and that the November IPO turns every institutional partner into a valuation argument. The solution: partner institutions should negotiate portability and exit clauses now, not in three years.

Mistake 2: Confusing Claude credits with computing capacity

Credits cover inference, not training or massive compute. A genomics lab that needs to process petabytes won't solve its problem with tokens. Credits are a lever for analysis and design — the computing power remains the responsibility of the lab or its partners, like SpaceX on Anthropic's side.

Mistake 3: Neglecting traceability of the models used

A result obtained with Claude Opus 4.6 is not comparable to a result obtained with Claude Opus 4.7 (Adaptive). Teams that fail to archive the exact model version, parameters, and execution dates will produce unverifiable results. The solution: systematically version API calls in lab notebooks.

❓ Frequently Asked Questions

Who are the 20 partners of the AI for Science program?

On the funders side: Alfred P. Sloan Foundation, Simons Foundation, Schmidt Sciences, Open Source for Science Fund, NSF and CAS. On the institutions side: Wellcome Sanger Institute, Francis Crick Institute, Imperial College London, EMBL, Gladstone Institutes, Space Telescope Science Institute. On the projects side: FlyWire, LEAP, AstroPath, gnomAD, Tau Ceti, Centre for Population Genomics, among others (October 2026).

Are the credits really free?

Yes, for researchers and teams selected through the program's three tracks, with no direct financial counterpart. But Anthropic sees it as an investment: the expansion announcement specifies that the program must exceed the initially planned 10,000 seats, and the premium seats at $15 per month show the monetization trajectory.

Why this announcement now?

Three combined reasons: capitalize on the week's momentum (drug discovery, GSA contract), fuel the growth narrative ahead of the November 2026 IPO under the ticker ANT, and lock in institutional partnerships that competitors won't be able to replicate quickly.

Are there independent alternatives for research?

Yes: DeepSeek V4 Pro (88) and Kimi K2.6 (84, available for self-hosting) offer solid performance, and GLM-5.1 (83) can be deployed internally. But none of these alternatives has a comparable network of institutional partnerships, nor the integration into funding channels that AI for Science represents.

✅ Conclusion

With this expansion to 20 institutional partners, Anthropic has moved from scientific patronage to infrastructure distribution — and on the eve of its IPO, that's exactly the story that needed to be told. All that remains is for laboratories to make the most of the credits without inheriting the dependency: open science deserves models whose lifecycle it controls. To see how Claude stacks up against its competitors in another area, check out our Claude vs ChatGPT comparison or our selection of the best LLMs for coding.