Anthropic’s researchers report mapping a hollow protein shell called a vault with Claude’s help. The result could give scientists another laboratory system for studying these particles, whose biological function remains unknown.
As of November 4, 2026, the company describes preliminary work, with a preprint submitted to bioRxiv and structures and further data still to come. The announcement date is unverified.
For readers assessing AI-discovery claims, the useful distinction is between structure, function and application.
A structure is not a biological explanation
Anthropic reports that L05, a protein from a phage—a virus that infects bacteria—assembled into a shell of 96 copies when produced in E. coli. Cryo-electron microscopy—reconstructing a three-dimensional shape from images of frozen particles—revealed its architecture. Its formation in native hosts remains unestablished.
The protein’s origin matters. A 2022 study had already produced vault-like particles in E. coli using the human version of the main vault protein. Making a human protein in bacteria is different from investigating a protein originating in a phage.
A 2023 preprint also identified possible relatives of vault proteins in phages through computational analysis. That established candidates for investigation, not experimental proof that they assembled into complete particles.
A useful way to judge the evidence is to ask three separate questions:
Structure: Do experiments show how the pieces fit together, rather than only predict a shape?
Function: Is there evidence linking the particle to an effect in its natural biological setting?
Application: Has a proposed use been demonstrated, rather than inferred from the particle’s appearance?
Progress on the first question can make the others easier to investigate. It does not answer them automatically.
The AI contribution needs its own evidence
Anthropic says humans directed the research while Claude assisted with database searches, DNA design and microscopy analysis. It reports 34 hours to an initial structural model and about a week to final three-dimensional maps, on internal infrastructure using hundreds of graphics processors. The account also describes a two-week protocol setback and an early map requiring correction.
The implication for evaluating AI assistance is that these timings do not isolate the model’s contribution from computing resources and scientific judgment. Nor are they a controlled comparison showing how much faster another laboratory would work.
For a lab considering a similar workflow, the useful comparison is time and cost to a checked result—not time to the first output. That means including expert review, correction effort and equipment, alongside the AI and computing bills. Initial models, final maps and whole projects are different endpoints.
The distinction between computational assistance and experimental proof also matters in RohitAI’s explainer on Antigravity’s chemistry tools.
Software access is not a complete research budget
As of November 4, Anthropic lists Claude Science as a beta application, not a separate model, available on Pro, Max, Ultra, Team and Enterprise. Ordinary Team and Enterprise access needs administrator enablement; the scientists’ Team plan enables it by default.
The usage documentation says Claude Science shares the plan’s usage limits, while connected services such as Modal bill users separately. Those terms do not establish an all-in price for reproducing the vault study.
There is a similar boundary around checking results. The documented background reviewer compares claims with execution records but does not rerun the analyses. A traceable result makes scrutiny possible; it is not automatically a correct scientific conclusion.
For a research team, a sensible first evaluation would be a familiar analysis with a reference result that an expert can assess. Keep the inputs, code and corrections, and compare the checked outcome, total effort and costs before expanding to an unfamiliar discovery problem.
For everyone following the science, the question to watch is what experiments can establish about the shell’s role—not simply how quickly a model of its shape can be produced.