Anthropic opened a public call on 22 October 2026 to bring "guest investigators" from the life sciences into joint projects with its own team. The format is short and bounded: eight to twelve weeks of direct collaboration, with the company picking up the cost of running its models and the scaffolding around them, plus a commitment to publish the results together and release the data and code that come out of it. The application window lasts roughly fourteen days from the announcement.
What was announced
The call is aimed at principal investigators and their groups at academic institutions, nonprofits and government laboratories. Anthropic lists five subject areas: causal human biology, molecular biology in both its structural and computational forms, medicinal chemistry, drug discovery and clinical science. Proposals go through a form linked from the official announcement.
What the company puts on the table has three parts. First, time from its own researchers working side by side on a real problem belonging to the applicant group, not on a demo case. Second, model and harness runs at no cost to the invited team — and that distinction matters: this is not just API access, it also covers the agent infrastructure Anthropic uses internally. Third, an agreed deliverable: analyses, artefacts and tools that are genuinely useful to the group's line of work.
On intellectual property the announcement is explicit. Guest investigators keep the rights to their data at every stage of the project, and also to any follow-on research they pursue independently off the back of the collaboration. Anthropic further stresses that it does not ask for the group's data either at application time or during the work, and in turn asks that application materials contain only already-public information.
The published selection criteria are two. One, that the applicant explains clearly what they gain from working directly with the company's researchers. Two, that they make a reasonable case that their output — a preliminary finding, an answer to an open question, a released dataset, a public resource — fits inside the eight-to-twelve-week window. The company concedes that properly validating most open problems takes far longer, and clarifies that what it is after is early evidence, not closed conclusions. Some projects may be extended beyond the initial period.
The calendar is tight at both ends: around fourteen days to apply, and roughly another fourteen from the closing date for selected participants to be notified by email.
Where it comes from
This is not the first piece on Anthropic's science board. The company itself places the call within a cluster of initiatives it says are deepening its presence in the life sciences and in research more broadly. It names two existing ones: a life sciences verification programme, meant to give scientists direct access to its most capable publicly available models for biology, and an AI-for-science initiative that hands out Claude credits to research groups.
In its own material, the company ties this announcement to work published a month earlier, on 23 September 2026, in which it claims Claude identified a novel enzyme system with CRISPR-like repeats. That is, in fact, the only concrete precedent the text puts forward.
What it means
For anyone who builds or maintains software, the interesting part is not the biology: it is the pattern, and what gets published afterwards.
The pattern is a vendor that, instead of selling licences and documentation, drops its own researchers inside the user's problem for a couple of months and pays for the compute. Anyone who has tried to get real value out of an agent system in a specialised domain knows where the bottleneck sits: not in the model, but in the scaffolding — how tools are chained, how each step is checked, where you stop. The fact that Anthropic explicitly includes its harnesses, not just its models, suggests the company knows where the friction actually is.
What remains afterwards is reusable material. The commitment to release data and code turns every project into a public worked example of orchestrating a model against a serious scientific problem. That tends to be worth more than any benchmark: it shows prompts, task decomposition and stopping criteria in a case somebody had to defend in front of reviewers.
Then there is the contractual side, which is the easiest bit to copy. Data rights that do not move, no dump of information to the vendor, an application built only from public information, and follow-on work that stays with the group. It is a sensible template for any collaboration with a model provider, scientific or not.
What the announcement does not say
There is no number of places. Nor any detail on how many Anthropic researchers are assigned per project, at what level of dedication, or what happens if a selected group reaches nothing presentable in twelve weeks.
The specific models and harnesses on offer are not named beyond the generic description. There is also no word on the licence under which code and datasets will be released, or on who decides what gets published when the result is thin or awkward. Groups keeping their data rights is not the same as having the final say over a joint publication.
The extension criteria stay open — "there may be an option" is not a commitment — and so does the programme's future: the company says it hopes to broaden the effort as it learns how best to support the scientific community, which leaves a second round unguaranteed.
Finally, the September enzyme result that serves as the calling card is a claim Anthropic makes about its own system. As with any finding published by the party selling the tool, what settles it is not the announcement but independent validation — and that, by definition, does not fit into eight weeks.