Claude and K Pro - cracking biology together

Following Anthropic’s announcements at The Briefing: AI for Science on June 30th, 2026, including Claude Science, Anthropic’s AI workbench for scientists, we are excited to share that our agentic platform, K Pro, currently reasons through frontier models, including Claude. We believe that together, K Pro and leading frontier models like Claude can be used together to solve biology to accelerate cures for many yet unsolved diseases.

Reasoning about biology is the bedrock of any AI scientist. Claude models can take a question about a disease mechanism, hold a dozen competing hypotheses at once, weigh the evidence, and propose the next experiment in seconds. Claude is among the most capable reasoning systems available, and we expect it to keep improving. That’s why we’ve currently made it our reasoning model of choice to power our AI scientist, K Pro - even as we build out our own biospecific models.

Reasoning alone is no longer the main bottleneck in modern biology, data and experimentation are. A model can propose a convincing explanation for why a tumour resists pembrolizumab through a particular pathway and still be wrong, because the answer only becomes clear when we examine what actually happens in patients and validate mechanisms experimentally at the bench. In practice, scientists generate hypotheses, design and run experiments, and then iteratively refine those hypotheses in response to the results.

That’s what we’re building with K Pro. K Pro pairs Claude's reasoning with the three things needed to run that loop that K Pro can provide:

‍

Expert skills

A general model knows a little about every method and the precise version of almost none. Real analysis runs on specific procedure: adjusting a survival model for tumour purity and batch effects before trusting a biomarker, or scoring a pathology slide the way a practising pathologist scores it rather than averaging the pixels. We encode those procedures as skills our AI scientist can call on, so a question about which patients respond to a PARP inhibitor is answered with the right covariates and controls, not a plausible-sounding summary.

‍

Patient data, analysed properly

You cannot reason your way to how a disease behaves in people; you have to observe it across many patients. Through MOSAIC and our wider network we work with 104 institutions and 156 leading clinicians on the largest spatial omics dataset in oncology. So when Claude proposes that a particular gene drives resistance, we can check it against thousands of real tumours: is that gene actually upregulated in the patients who relapsed? Does the signal survive once you adjust for stage and line of treatment? The hypothesis either holds across real cohorts or it does not.

‍

The wet lab

Some questions have no existing answer because the experiment has never been run. There, reasoning hands off to the bench. A target K Pro flags as promising can go straight into a knockout cell line or a patient-derived organoid assay. If knocking out the gene kills the resistant cells, the prediction earns its place. If it does not, we have saved a programme from chasing it.

‍

‍

Put the three together and each cycle gets less wrong, with the corrections coming from patients and experiments rather than from more text. This is what we mean by biological superintelligence. It is not a bigger model; a bigger model is just a better reasoner about the same incomplete picture. It is the connection between the reasoning and the biology it reasons about, and that connection is hard to build. Almost nobody has spent ten years assembling the patient network, the clinical relationships, and the lab capacity to tell that model, reliably and across thousands of cases, when it is wrong.

So we want the frontier models we use to keep getting better. That’s why we’re building MCP connectors - in January 2026, we announced the inclusion of our Pathology Explorer model in Claude for Life Sciences via MCP. The sharper the reasoner, the more valuable the loop around it, because the hypotheses are tighter and the experiments more pointed. Our job is to keep K Pro grounded in real patients and proven at the bench. If frontier reasoning keeps advancing and we keep deepening that grounding, the two together could move biology further in the next few years than either could alone.