Announcements from AI companies tend to draw attention to the performance figures of new models. But this Anthropic announcement contains no new model. What emerged were two things: a workbench for scientists and an in-house program to search for drug candidates. The company is competing not on raw intelligence but on how that intelligence is put to use. This organizational decision is the point worth reading.
01What Happened — Simultaneous Launch of a Workbench and a Drug Discovery Program
On July 1, 2026, Anthropic formally launched "Claude Science" for researchers and the pharmaceutical industry. TechCrunch called this a strategy of differentiating on the workflow (= the flow of work) side rather than on a new model. At the same time, CNBC, Reuters, and Endpoints News reported the start of an in-house drug discovery program in the biopharma field. Then on July 4, it was reported that Anthropic is considering entering the drug development business (The Verge). Within a few days, two moves appeared side by side: providing tools, and the possibility of entering as a direct participant. Its competitor OpenAI announced "GeneBench-Pro," an evaluation benchmark for the bio and science domain (= a yardstick for measuring an AI's biological capabilities), in the same period, and science is becoming the next main battleground.
02The Backstory — The Path Laid by Insilico and Moderna
This move did not come out of nowhere. AI drug discovery already has an accumulating track record. Rentosertib, from Insilico Medicine (a company name derived from "in silico" (on a computer)), is a treatment for idiopathic pulmonary fibrosis (IPF = an intractable disease in which the lungs gradually harden) in which AI handled everything from the search for the target (= the molecule in the body that the drug acts on) to the design of candidate molecules. It passed Phase IIa (= the mid-stage clinical trial that first confirms efficacy in patients) in early 2026. It is regarded as one of the early examples of a molecule in which AI was involved from search through design to reach this stage. It is now entering preparation for Phase III (= the large-scale trial before approval). Moderna, meanwhile, has partnered with OpenAI company-wide since 2024 and is known as an early case of deploying AI fully across internal operations. AI that handles target search, and a use that dissolves AI into the entire corporate operation — these two currents form the groundwork of the current phase.
03What the Numbers Show About Where We Stand, and How to Read Them
AI drug discovery is no longer an exceptional attempt. More than 173 AI-derived candidates are said to be in clinical trials (though note that the definition of "AI-derived" — which stage of the process AI was involved in — is not consistent). The pass rate for early Phase I (= the first trial in which a drug is given to humans for the first time) is reported to reach 80–90%, exceeding the conventional roughly 52%. Still, these figures need to be read carefully. The sample used for the tally is small, and the picture changes depending on which phases are compared. In addition, the real gateway in drug discovery lies not in Phase I, which examines safety, but in Phase II, which tests efficacy. It is reasonable to view the track record of AI-designed molecules clearly outperforming the conventional in Phase II as still limited. The high pass rate is confounded not only by AI's design capability but also by a conservative choice of targets that makes it easy to pick already-validated ones. The contribution specific to AI needs to be separated out from that. The technical base has certainly expanded. AlphaFold3 (a protein structure prediction AI developed by Google DeepMind and Isomorphic Labs) greatly sped up the prediction of binding between a drug and its target. The ways companies connect also vary. There are inter-company partnerships such as Bayer × Cradle (an AI company) and Eli Lilly × NVIDIA, and there are also moves of a different nature from partnership, such as Pfizer taking in Boltz, a structure prediction model described as open source. In 2025, Recursion acquired Exscientia. But this consolidation reads less as evidence of momentum than as industry restructuring and rescue under clinical failures and funding difficulty. It can also be read as a sign of the difficulty of making AI drug discovery viable as a standalone business.
04Why Now — Structural Pressure and the Board's Move
The pharmaceutical industry is under strong pressure. The cost of developing a single new drug has swelled over 20 years from $1 billion to $2–3 billion. On top of that, 2025–2030 coincides with the patent cliff (= the period when the patents of leading drugs expire in concentration and sales drop sharply). Costs are rising, and the breadwinners are being lost. That said, what AI can shorten is mainly the search and preclinical stages. What accounts for the bulk of development costs is large-scale clinical trials, and AI does not act directly on those. Its immediate effect on total development costs should be seen as limited. There is also a mismatch in the time axis. A candidate found with AI today will reach approval, at the earliest, in the second half of the next decade or later, and does not directly rescue the imminent 2025–2030 cliff. The substance of the expectation lies in reinforcing the next-generation pipeline (= the group of drugs under development). Here Anthropic's own move overlaps. In April 2026, Novartis CEO Vas Narasimhan (a physician by background, involved in developing more than 35 new drugs) became an Anthropic board member. The appointing body was the Long-Term Benefit Trust (LTBT = Anthropic's own oversight body that prioritizes long-term benefit to humanity over shareholder interests). This Trust holds the authority to nominate board members, and with this appointment, the side nominated by the Trust exceeded a majority of the board. Detailed coverage is left to a prior report (/journal/novartis-anthropic-board.html), but the significance of bringing a drug discovery participant into the core of the governance structure is not small.
05How to Read Anthropic's Strategy
From here on is interpretation. First, Anthropic appears poised to differentiate not on the performance of a new model but on workflow. That Claude Science came without a new model is likely a decision to win on how well it fits into the research site rather than on the absolute value of intelligence. General-purpose models tend to end up on par with competitors, but a workbench closely attached to research work creates switching costs and builds stickiness. Second, the start of the in-house drug discovery program can be read as a step from tool provider toward participant in drug discovery. If it only sells tools, the results belong to the customer, and AI's degree of contribution is hard to see. By producing candidates itself, it can show how far AI can go as its own track record. However, the drug development business itself is reported only to be "under consideration." This too can be read as a sign that Anthropic itself is gauging the distance at which it might compete with its pharmaceutical customers. Third, the LTBT's nomination of Narasimhan can be seen as groundwork that supports this direction from the governance side. For a Trust that upholds long-term benefit over shareholder interests to place someone with drug discovery experience on the board is a structure that makes it easier to hold up new drug development as a benefit to humanity rather than short-term sales. For the design philosophy of the entire product line, see also the prior report (/journal/claude-science-pharma-future.html).
06Implications for the Pharmaceutical Industry and the Front Line
Looking ahead, the boundary between those who hold AI and those who hold drugs will likely dissolve further. If the flow from partnership to direct participation advances, then for pharmaceutical companies an AI company becomes both a seller of tools and a potential competitor. In this phase, what the pharmaceutical side will first brace for is, I think, several practical issues. One is the ownership of IP (= intellectual property). Who owns the rights to molecules generated by AI, and what happens to sovereignty over the data used for training, are areas where contracts and institutions remain undeveloped. The second is the stance of regulators. How the FDA (US Food and Drug Administration) and the PMDA (Japan's Pharmaceuticals and Medical Devices Agency) will evaluate AI-based development is not yet settled. The third is a chilling of data sharing that arises as AI companies become competitors. Detailed figures are compiled in a prior report (/journal/ai-drug-discovery-2026.html). Such changes reach the work of the front line as well. In material review (= the task of confirming that promotional materials comply with regulations), new judgments are required about how to verify the evidence data and design process in which AI was involved, and how much may be documented. In medical (= the department responsible for providing medical information), the responsibility to accurately relay to healthcare professionals the mechanism of action and trial results of AI-derived candidate drugs grows heavier. In the domain of safety (= the monitoring of side effects), the eye for whether an AI-designed drug behaves in unexpected ways will be called upon all the more. The smarter the tool becomes, the verification responsibility of the humans who handle it does not lighten. If anything, it grows heavier. This, I believe, is the core of how the front line should take in AI drug discovery in the new era.