Claude Science: Anthropic steps into the lab, where AI can change more lives but the obstacles are radically different

By Momentum IA · June 30, 2026.

Anthropic introduced Claude Science, a unified workspace designed for scientific researchers. The first thing to understand —and which the company itself stresses— is that **it is not a model**: it is an orchestration layer that consolidates scattered tools and databases into a single environment preconfigured for genomics, single-cell analysis, structural biology and cheminformatics. The system renders protein structures in 3D, generates publication-ready figures and includes the complete code, environment details and creation history of every result. It runs on Opus 4.8 with no special access, integrates with Nvidia's BioNeMo Agent Toolkit and supports the Model Context Protocol for custom extensions.

This is the latest stop on a tour of verticals that Anthropic is making at a brisk pace: from AI for code to law and finance, from there to cybersecurity and now to science. It does not come out of nowhere: both Fable 5 and Mythos 5 already incorporate capabilities in molecular biology and genomics. Claude Science is, in this sense, the crystallization of those technical bets into an interface accessible to anyone who does not want to build pipelines from scratch.

**Why the choice of «workbench» matters more than it seems**

The decision not to present this as a model specialized in science is, in our view, one of the smartest aspects of the proposal. Chirag Shah, a professor at the University of Washington, explains it well: previous attempts to take a foundation model and fine-tune it over a domain like biology «have not worked particularly well». The workbench approach sidesteps that pitfall: instead of retraining for a discipline, it orchestrates already-proven capabilities around the scientist's real workflows —hypothesis formulation, literature review, data analysis, visualization—. It is a lesson in maturity: AI does not have to be a biochemist; it has to be an infrastructure that amplifies the person who is one.

That also means the proposal is not unique. OpenAI has the FrontierScience benchmark and Google offers Gemini for Science. Claude Science does not invent the category; it aims to win on convenience and depth of integration. In the enterprise market, that is no small thing: whoever reduces day-one friction the most wins the team's inertia.

**Science's obstacles are qualitatively different from those of code**

AI has shown extraordinary gains in programming because code is an environment of immediate, almost infinite feedback: you run, you fail, you fix, in milliseconds. Science does not work that way. John Thickstun, of Cornell, sums it up precisely: «With many scientific applications, you are lucky if you get feedback once a year». A structural biology experiment can take months; a genomics replication, weeks; the validation of a clinical finding, years.

This does not invalidate the tool —accelerating the analytical and literature-synthesis side is already a huge advance—, but it does narrow expectations. Claude Science can do very well what happens *before* the experiment (design, literature review, hypotheses) and *after* (data analysis, visualization, writing). The physical experimental loop remains a human domain, and probably will for years. Anyone who sells this as «the AI that will discover the cure for cancer on its own» distorts reality; anyone who sees it as a multiplier of researchers' time speaks more honestly.

**Our take: the most important bet, even if not yet the most profitable**

The underlying thesis of Momentum IA is that AI's transformative value in the long term lies not in corporate chatbots or coding assistants —which are today's business— but precisely here: in the acceleration of scientific discovery that could lead to eradicating diseases, extending health and, ultimately, reshaping what the human condition means. Claude Science is a step in that direction.

That Anthropic devotes resources to a «harder» vertical —as Thickstun notes— when it could concentrate on the already profitable ones says something about the company's culture. It is not philanthropy: science is a huge market with high institutional willingness to pay. But it does imply a longer time horizon and a tolerance for failure that quarterly cycles typically punish.

As sector context, the race toward scientific applications is accelerating in parallel across the major labs. What is most relevant is not who arrived first (Google has spent years at DeepMind with AlphaFold, which demonstrated the potential emphatically), but who builds the usage layer that connects the model with the researcher's real workflow. That is the battle Claude Science aims to win, and one in which integration with Nvidia BioNeMo, MCP support and the curation of the environment are more decisive than the raw power of the underlying model.

In the short term, the impact will be marginal and sector-specific: research teams that adopt the tool will gain efficiency in analysis and synthesis, but there will be no quantum leaps in discoveries overnight. In the medium and long term, if the feedback loops compress —with faster synthesis, more precise experimental design, smoother global collaboration—, we are looking at one of the most powerful levers humanity has ever built to tackle its hardest problems. The road is long and the obstacles are real. But the direction is what matters.