HomeArticle

Dario's most anticipated AI mission, a retrospective of Anthropic's biomedical business layout

海外独角兽2026-11-18 08:56
How Frontier Lab reshapes the landscape and value distribution of the pharmaceutical industry

01.Key Takeaways

⁠•Biomedicine is the next high-value direction that Anthropic is concentrating heavy investment in after Coding. Drug R&D itself requires top-tier intelligence, and every key progress in the drug R&D process can generate considerable asset value, making it possible for AI's charging model to shift from API billing to result-based billing.

⁠•The current priority of Anthropic is to close the loop as soon as possible. As computing power accelerates the small loop of molecular design to run through faster, the bottleneck shifts more to wet lab experiments. Molecular design can already be generated quickly, so the fastest area for Anthropic to close the loop is to use Claude to call molecular design models such as BoltzGen, and accelerate the efficiency of models on GPU. At the same time, molecular verification is still restricted by wet experiments. Anthropic has also started to build its own laboratories and actively cooperate with external laboratories to obtain reliable feedback loops faster and promote research.

⁠•We believe Anthropic is the strongest and most serious key player in the AI for science direction. Compared with OpenAI, Anthropic is more focused. Compared with Eli Lilly, Anthropic is more AI-Native. Although OpenAI started early in the biological field, its layout spans personal health, medical services and life sciences, which is less focused than Anthropic. As an example of a pharmaceutical company deploying AI drug development, Eli Lilly owns its own data and pipelines, and mainly deploys AI by building platforms with partners. At present, its pace and organizational form are still relatively conservative.

⁠•Anthropic is gradually participating in pharmaceutical development, but considering its relationship with pharmaceutical companies, it currently focuses on pre-clinical and rare diseases. Anthropic is advancing its own pipelines to understand the role and limitations of AI models in the real R&D process, and it does not rule out the possibility of entering the pharmaceutical industry in the long run. But at this stage, the relationship between them and pharmaceutical companies needs to be balanced. Since they cannot compete with their large pharmaceutical enterprise customers, Anthropic limits its existing R&D to before human trials, and focuses on rare diseases that large pharmaceutical companies do not invest many resources in.

⁠•The improvement of Claude's capabilities is also changing the form of Biotech organizations, making "pipeline in a person" possible. In the future, a Biotech will most likely be an experienced leader, with the help of a large number of AI Agents, cooperating with external laboratories to complete the drug R&D and management work that used to require a huge team. Mithril, which Anthropic acquired this year, provides a good example.

⁠•Anthropic will change the industry pattern and value distribution of the pharmaceutical field, and eventually their positioning may be the TSMC in the pharmaceutical industry: Anthropic is responsible for early drug discovery and R&D, while pharmaceutical companies are responsible for late-stage clinical trials and sales. In the first half of this year, they provided an entry for scientists to work through Claude Science, and unified the scientific laboratory ecosystem through MHS. These layouts on the scientific research ecosystem are the beginning of their pattern change in this field. In the second half of the year, they began to accelerate protein design, support one-person pipelines, and go deep into the biopharmaceutical workflow, which made us start to see signals of changes in the value distribution.

⁠02.Why Anthropic Attaches So Much Importance to Biomedicine

⁠Compared with OpenAI, Anthropic's betting style is more focused. The previous all-in on Coding brought huge success to the company, allowing Anthropic to achieve a comeback against OpenAI. Now after Coding, biomedicine has become the next high-value breakthrough paradigm. It is also clearly written in Anthropic's S1: Agents oriented to scientific discovery will gradually become a reality in 2026. The company puts life sciences and robotics side by side in the position of the most worthy new business to invest in. Dario himself also said in Anthropic's roadshow that he believes diseases including cancer, Alzheimer's disease and heart disease will become a thing of the past within ten years.

Anthropic's roadshow slides list Life Sciences as one of the two major new directions in the future

⁠We believe that Anthropic attaches so much importance to biomedicine, mainly for several reasons:

⁠Biomedicine may be the most promising field for AI to explore result-based payment.

⁠Anthropic clearly stated in its prospectus that its next step is to become a virtual collaborator, with the core capability of directly delivering task results. Among many tracks, biomedicine is the best model for result-based payment. Every key progress in the drug R&D process can form a tradable asset value: as candidate molecules reach the agreed standards and clinical trials achieve positive results, partners can obtain considerable milestone payments accordingly.

⁠Drug discovery is the next high-value task after AI Coding, and it is also the jewel in the crown of intelligence.

⁠In 2026, global annual pharmaceutical expenditure exceeds 2 trillion US dollars, and a successful drug can bring considerable commercial returns. Taking GLP-1 drugs as an example, the annualized sales of the main products of Eli Lilly and Novo Nordisk in the first three quarters of this year are close to 100 billion US dollars. The attractiveness of drug discovery is not limited to the market size: it requires understanding complex biological mechanisms, putting forward hypotheses in incomplete information, and continuously correcting judgments through experiments, which puts forward extremely high requirements for cutting-edge intelligence.

⁠Dario himself has a strong interest in biology, and biology is one of the few bets Anthropic has made.

⁠Before devoting himself to AI, Dario was engaged in biophysics research and did postdoctoral research at Stanford University School of Medicine. His father died of hepatitis in his early years, and less than five years after his death, medical progress reduced the fatality rate of this disease from 50% to 5%. This incident shaped Dario's view of technology and time. In Machines of Loving Grace, he listed biology and health as the top of the AI application directions he most looks forward to: the AI he envisions is a "virtual biologist" that can put forward hypotheses, design experiments, and invent research methods. Today's Anthropic is gradually approaching this goal step by step.

⁠03.Core Objectives

⁠Anthropic hopes to replicate the leap of Coding in life sciences: from being an assistant to scientists to completing research tasks end-to-end. Eric Kauderer-Abrams, Head of Life Sciences, said in a public interview that he hopes scientists can get the experience that software engineers already have: someone to discuss with, and tasks can be delegated. He described the goal as a "research assistant that exceeds human capabilities".

⁠Behind this goal, there are two main lines for the life sciences team:

Accelerating Scientific Discovery as a Tool

Claude Science: AI Research Assistant

⁠In June 2026, Anthropic launched Claude Science, integrating literature retrieval, data analysis and paper writing into the same scientific research workbench, allowing scientists to delegate multi-step research tasks.

⁠The company juxtaposes Claude Science with Claude Code and Claude Cowork, corresponding to scientific research, software development and knowledge work respectively. It supports the display and interaction of various scientific results; retains analysis history; schedules computing tasks; connects to scientific research databases and professional tools; allows users to expand according to their own research fields.

⁠With built-in Skills and connectors, the product can use more than 60 databases and scientific resources, covering proteomics, structural biology, chemistry, genomics and literature research. Users can add Skills through conversation, or write and upload ZIP files by themselves, or import from GitHub. Existing local and remote MCP services can also be accessed.

⁠MHS: Claude Enters the Lab Bench

⁠If Claude Science carries the workflow of researchers in software, then MHS (Model Hardware Standard) is an attempt by Anthropic to extend AI to laboratory equipment and physical operations.

⁠On August 27, 2026, Anthropic opened the MHS research preview. This specification, which originated from its cooperation with non-profit organizations, tries to provide a common interface for different devices, so that AI can discover instruments, read status, call operations, and understand the functions and safety restrictions of the equipment. It is not bound to the model, and can also be used by agents through methods such as MCP.

⁠At present, a major bottleneck of automated laboratories lies in the integration cost. The automation degree of a single instrument has been improving in the past, but how to combine them together has become a new problem. An experiment usually requires the cooperation of multiple instruments. If instruments can describe their capabilities in a common way, the experimental software will have the opportunity to call the equipment according to the task, transmit the execution results back to the model, and then decide the next step.

⁠This is also a key step towards the "unmanned laboratory", reducing errors caused by manual operations and improving experimental reproducibility. Claude can obtain as much information of the instrument as possible through MHS for analysis. It can truly close the loop in the laboratory and shorten the feedback loop.

⁠We believe that this is very likely to become a unified standard for laboratory hardware to access AI tools, similar to AWS in the laboratory. In the future, researchers will have the opportunity to call experimental capabilities through software, instead of building their own laboratories and connecting instruments one by one for each study. Links such as pipetting, culturing, imaging and analysis may gradually become composable experimental services.

⁠At present, Anthropic has announced the first batch of cooperation cases:

⁠Developing Drugs in Person but Stopping at Pre-clinical Stage, Focusing on Rare Diseases

⁠On the other hand, Anthropic is participating in drug discovery, but it has also drawn boundaries for its own R&D activities. Eric Kauderer-Abrams, Head of Life Sciences, said in an interview that the company will carry out early drug discovery. Only by participating in drug R&D in person can we truly understand where the models and products fail. By taking on projects directly, the company can get closer feedback instead of only relying on customers to describe problems. The boundary here is before the start of Phase I human trials.

⁠In terms of disease selection, Anthropic focuses more on indications that traditional commercial institutions lack motivation to invest in. Eric linked this choice to the company's PBC (public benefit corporation) positioning: avoiding projects that customers find attractive, and investing in studies that have therapeutic value but may not be carried out by others. This not only provides a mission-level reason for the company to participate in R&D in person, but also helps reduce conflicts of interest with pharmaceutical customers.

⁠We speculate that this arrangement may serve two strategic purposes at the same time.

⁠1.Concentrate resources on R&D links where AI has more opportunities to accelerate, and explore the organizational form of "super biotech". Early drug discovery involves a large number of tasks that can be assisted or gradually undertaken by AI. Anthropic's personal participation in these links can obtain a large amount of data in real projects, and close the loop to improve model capabilities. It also explores using a smaller team to promote more R&D projects. In contrast, the steps after human trials rely on the execution capabilities accumulated by pharmaceutical companies for a long time. Stopping its own R&D before clinical trials may mean that Anthropic hopes to concentrate on investing in the links where it has more advantages, and then let other institutions take over the subsequent development.

⁠2.Clarify the business boundary with customers and strive to become the TSMC of the pharmaceutical industry. If pharmaceutical companies want Claude to go deep into the R&D process, they need to trust that Claude will not become a direct competitor of the core pipeline at the same time. From this perspective, TSMC may be a reference: by focusing on wafer manufacturing and emphasizing no competition with customers, it supports chip design companies to hand over key production links to themselves. If Anthropic can establish similar trust, more pharmaceutical companies will be willing to adjust their R&D workflow around Claude.

⁠Participating in Cutting-edge Scientific Discovery

⁠In the spring of 2026, the company established an internal basic biology research group, bringing model training, computational analysis and wet experiments into the same team. On September 18, Eric Kauderer confirmed to Reuters that the company was already operating a wet laboratory. He believes that biological research must eventually be tested by real experiments, and some work will be done faster by themselves, and first-hand research experience can be obtained. Therefore, the company adopts a method similar to most Biotechs: some experiments are completed in its own facilities, and work suitable for outsourcing is handed over to external cooperative institutions.

⁠In September 2026, Anthropic announced its internal life sciences research team and laboratory, and disclosed an early discovery: after researchers gave the task of finding a novel reverse transcriptase system, about 950 Claude agents ran for 21 hours and identified a system with special repeating sequences, which the team named ART.

⁠In November 2026, Anthropic demonstrated the ability of the new model to solve complex biological computing problems. In the AAV capsid structure prediction task, Mythos 5.2, based on two open-source models OpenFold3 and ESMFold2, developed a method to predict the complete structure using the symmetry of the protein shell. The AAV capsid is composed of dozens of protein chains and is an important delivery vector for gene therapy. Complete structure prediction usually requires a large amount of computing resources. According to the company's disclosure, the method developed by Claude reached the preset structural accuracy standard for 28 of the 34 natural or engineered capsids tested, demonstrating the potential of the model to expand the capabilities of open-source scientific research tools by writing code and improving computing methods.

⁠But at this stage, all experimental operations are completed by human scientists. Although Anthropic is trying automation solutions such as MHS, the molecular biology research of this team often needs to temporarily adjust the experimental process, which is not very suitable for this automation method at this stage.

⁠Acquiring Biotech and Its Pipelines

⁠In April 2026, Anthropic completed the acquisition of Coefficient Bio, focusing on supplementing the team and technical capabilities of AI and biomedicine. This is a typical talent acquisition. CEO Aris Theologis became Anthropic Lead of Life Sciences, and CTO Nathan Frey joined Anthropic to lead drug discovery and development research.

⁠In July 2026, Anthropic acquired Mithril Medicine. Mithril develops bispecific antibodies, trying to repair the connection between muscle fibers and the surrounding supporting structures. Its research directions include advancing two early bispecific antibody R&D pipelines, targeting LAMA2-related congenital muscular dystrophy and α-dystroglycanopathy respectively. This is a one-person Biotech company. Founder Cody Tranbarger confirmed that these projects became Anthropic's first internal drug R&D pipelines, and he joined the company to continue promoting the projects and lead broader internal drug development work.

04.Key Personnel