Anthropic has opened a new route for academic and nonprofit scientists to work with its own researchers. In an open call for guest investigators in the life sciences published October 22, 2026, the company invites principal investigators to run special projects using Claude on key scientific questions, working directly with Anthropic staff for 8 to 12 weeks. Anthropic says it covers the compute, investigators keep rights to their data, and the results are published together with open-sourced data and code.
If you run a lab, the practical question is whether this is a grant, a service or a partnership. It is closest to a time-boxed partnership. This post lays out the terms, the selection criteria, what is still unclear, and how the program sits next to Anthropic's other science offers.
Guest Investigator Program at a glance
| Question | Answer (per Anthropic) |
|---|---|
| Who runs it? | Anthropic researchers, working directly with guests |
| Who can apply? | PIs and research groups at academic institutions, nonprofits or government labs |
| How long? | 8 to 12 weeks, some projects may be extended |
| Focus areas | Causal human biology, structural and computational molecular biology, medicinal chemistry, drug discovery, clinical science |
| Cost to investigators | None; Anthropic runs its models and harnesses |
| Data rights | Investigators retain rights at every step; Anthropic does not ask for data |
| Outputs | Published together; data and code open-sourced |
| Application window | About 14 days; no closing date stated on the page |
| Notification | By email about 14 days after applications close |
What guest investigators actually do
The announcement describes projects in which scientists use Claude on questions that matter to their own research, with Anthropic researchers alongside. The expected outputs are analyses, artifacts and tools. Anthropic's examples of what a good proposal might yield are preliminary discoveries, answers to open questions, open datasets or public resources.
The key phrase is "run its models and harnesses." A harness is the scaffolding around a model: tools, code execution, retrieval, long-running loops. Anthropic is offering not just model access but its own agent setups, tuned by the people who build them. That matters because results in science depend heavily on harness design, as we saw in Opus 5.5 agents proposing room-temperature magnetic semiconductor candidates and in the Claude quantum circuits research.
A folded protein ribbon, standing in for the structural biology projects a guest investigator could run with Claude
Terms that matter to a lab
Three commitments stand out, all stated on Anthropic's page.
- Data stays with you. "Guest investigators retain the rights to their data at every step during their project," and Anthropic says it does not ask for data in the application or during the project. Application materials should contain only publicly available information.
- Follow-on work is yours. Investigators keep rights to any independent follow-on research that comes out of the collaboration.
- Outputs are open. Results are published jointly and the associated data and code are open-sourced. If you need to keep a result proprietary, this is not the program for it.
The open-sourcing requirement is the real trade. You get free compute and expert collaborators, and the community gets the artifacts.
How selection works
Anthropic lists two criteria. Applicants must explain clearly how they would benefit from working directly with Anthropic's researchers, and they must make a reasonable case that useful outputs can be produced in 8 to 12 weeks. Preliminary evidence counts: Anthropic notes that full validation of most open problems will take longer than the project window.
That framing is honest about what a quarter of work can deliver. A strong proposal is therefore a narrow question with a computational or analytic route, not a full drug-discovery campaign. Wet-lab validation is a follow-on, not the deliverable.
What is not specified
Several details are absent from the announcement, and readers should not assume answers.
- Number of slots. The page does not say how many investigators will be chosen.
- Closing date. It says applications are open for about 14 days but gives no date.
- Model access. It does not name which models or tiers guests will use.
- Compute limits. "No cost" is stated, but caps are not.
- Geography. No country restrictions are mentioned either way.
Check the application page before drafting a proposal.
How it fits Anthropic's other science programs
Anthropic now runs several overlapping efforts, and they serve different needs.
| Program | What it offers | Best for |
|---|---|---|
| AI for Science | API credits for researchers at research institutions, launched May 2025; the page gives no amount | Labs that want to build their own pipelines |
| Life Sciences Verification Program | Vetted access to Anthropic's most capable public biology models with tailored safeguards | Biotechs and labs needing broader capability |
| Guest Investigator Program | 8 to 12 weeks of direct collaboration with Anthropic researchers | PIs who want hands-on help with one question |
| Federal credits | Claude credits for government science agencies | Public-sector labs |
For the verification track, see our coverage of the Life Sciences Verification Program and FlashPairformer, where Anthropic said Claude sped up more than 30 biomolecular models roughly 4x. For the public-sector side, see Anthropic's $150 million Genesis Mission credits.
The competitive picture
Anthropic is not alone in courting scientists. OpenAI has a life-sciences workspace, covered in OpenAI Rosalind Workbench, and robotics-driven wet-lab startups are emerging, as in Warp Sciences' plan for robot-run drug-discovery labs. The common pattern is labs competing for scientists' attention with free access, because published results are the best marketing for a model's scientific ability.
That is not cynical, just useful context. The guest model puts Anthropic's own researchers on the line for the outcome, which is a stronger commitment than credits alone. It also gives Anthropic real-world evaluation data, in the form of published projects, that benchmarks cannot supply. For a broader view of where language models actually help in math and disease research, see how language models solve math and disease problems.
Should you apply? A practical checklist
- Is your question computational or analytic? Eight to twelve weeks suits structure analysis, literature-scale data mining, assay design, or reanalysis of public data.
- Can you publish openly? The code and data will be open-sourced.
- Do you have a public-information version of your pitch? The application must not contain private data.
- Can you name what Anthropic's researchers add? The first criterion is explicitly about the benefit of direct collaboration, so say what you cannot do alone.
- Can you show a result in the window? Aim for a preliminary finding, dataset or tool, not full validation.
If your work involves higher-risk biology, the safeguards discussed around the Verification Program apply to the field generally, so expect scrutiny of dual-use topics.
What strong proposals might look like
Anthropic's page names five focus areas, and each maps to a different kind of 8 to 12 week project. These are our illustrations of the shape, not examples Anthropic gave.
- Causal human biology. Reanalysing public genetic and expression datasets to rank candidate causal genes for a disease, with the ranking and code released as an open resource. The deliverable is a prioritised list and a reproducible method, not a validated target.
- Structural and computational molecular biology. Running and comparing structure-prediction models on a family of proteins a lab cares about, then publishing the benchmark set. The FlashPairformer work shows why compute cost, not ideas, often limits this kind of survey.
- Medicinal chemistry. Mining published structure-activity data to propose analog series, with the model's reasoning logged so chemists can audit why each compound was suggested.
- Drug discovery. Building an open triage pipeline that screens public compound libraries against a target model and documents false-positive rates.
- Clinical science. Extracting structured outcomes from published trial reports into an open dataset, with human spot checks measuring extraction error.
In every case the output is something others can reuse, which fits the open-sourcing condition. Proposals that depend on private patient data, proprietary compound libraries or wet-lab results will struggle, since Anthropic asks for public information only and does not take data during the project.
Risks and limits to keep in mind
Free compute and expert help do not remove the usual problems of AI in science. Model-generated hypotheses can be fluent and wrong, so every claim needs an independent check, ideally in the lab that proposed it. Eight to twelve weeks is short, so scope creep is the main practical risk. And because outputs are published jointly, authorship and credit should be settled early; Anthropic's page does not describe how credit is assigned, so raise it in your application conversation rather than assuming.
There is also a reproducibility angle. Because Anthropic says it will open-source the accompanying data and code, outside groups can rerun the analysis. That is a stronger standard than most vendor science claims meet, and it is the best reason to take the published outcomes seriously when they arrive.
What to watch next
Anthropic says it may expand the effort in the future. The signals to track are the first published joint outputs, how many projects extend beyond 12 weeks, and whether results include real open datasets rather than demonstrations. Until those appear, the program is a promise with clear terms rather than a proven pipeline.
Related reading
- Anthropic Life Sciences Verification Program and FlashPairformer
- Anthropic $150M Genesis Mission credits
- OpenAI Rosalind Workbench
- Opus 5.5 agents and magnetic semiconductor candidates
- Claude quantum circuits research
- Warp Sciences robot wet labs
- How language models solve math and disease
- Official: Anthropic's announcement
Terms are as stated by Anthropic on October 22, 2026 and may change.