The discovery bottleneck

On the thirtieth of June, Anthropic launched Claude Science, a research workbench for life sciences built around a coordinating agent, more than sixty curated skills, and a second model whose job is to check the first one's citations and calculations. Alongside it, the company announced its own drug discovery programme, aimed at neglected diseases, on the stated logic that those are the indications traditional biopharmaceutical companies would not consider attractive.

The product is interesting. The move underneath it is more interesting, because it reopens a question that has been asked badly for a decade. Can a company with no pharmaceutical history discover drugs, and if so, what exactly is it that it has, and what is it that it lacks.

The advantage is not intelligence, it is the cost of being wrong

Incumbent discovery is not slow because the people are slow. It is slow because it is organised around a portfolio logic that made sense when a single asset cost a billion dollars and took fifteen years to reach a market. Every gate, every committee, every kill decision in that system exists to protect the bet. The apparatus is expensive, and it is expensive on purpose.

An entrant that can compress hypothesis generation does not inherit the apparatus. That, and not model quality, is the whole of the advantage. It is not that the entrant is smarter, it is that the entrant's cost of being wrong is an order of magnitude lower, and an organisation with a low cost of being wrong can afford to be wrong far more often. Which, in discovery, is the same thing as being fast.

Three consequences follow.

There is no sunk-cost defence of a legacy platform. Incumbents hold chemistry stacks, screening infrastructure, and twenty-year careers built around a particular way of finding things. That capital argues for itself in every meeting, and it argues persuasively, because it was expensive and it once worked.

Portfolio economics start to permit small markets. If discovery cost falls by an order of magnitude, a forty-thousand-patient indication becomes something other than charity. Anthropic's neglected-disease framing is the tell here. It is the one part of the market where the company can operate without competing directly against the customers it has just sold a workbench to, and it is also, coincidentally or not, the part of the market where an entrant's economics actually work.

The talent is not pre-shaped. A team assembled from computational biology and software rather than from Big Pharma R&D will ask questions the incumbent has already ruled out. Some of those were ruled out for excellent reasons. Some were ruled out in 2009.

What the entrant does not have

Discovery is the cheap part. This is the flaw in almost every technology-led entry into the sector, and it does not improve with repetition. The bottleneck in bringing a medicine to patients is not finding candidates, it is clinical development, regulatory strategy, manufacturing at scale, and reimbursement. Compressing the front of the pipe does not remove the constraint, it widens the mismatch. An entrant holding fifty promising leads and no clinical organisation holds an expensive filing cabinet.

Tacit knowledge does not appear in the training corpus. Why a particular scaffold failed in toxicology and was never written up. Why a target class is politely avoided at conferences. Which assay lies, and in which direction. That knowledge lives in people who have been burned by it, and it is not in the literature, because failure is not published. The entrant has not been burned yet, and the model has not read the thing that was never written.

Regulators do not reward novelty of method. The FDA and the EMA evaluate evidence, not elegance. An entrant arriving with an unfamiliar provenance story for its data carries a higher explanatory burden than an incumbent, not a lower one. The absence of institutional history is a liability in the room where it counts.

And corporate patience, which reads as groupthink from outside, is also risk absorption. A large pharmaceutical company can carry a programme through three bad years because the rest of the portfolio pays for it. Independent entrants run out of money at precisely the point where the science stops being cheap.

Hallucination, and where the human actually sits

The question is usually put as whether the model makes things up. It does, and that framing gets you nowhere, because it invites a blanket answer to a problem that is not blanket-shaped.

The useful question is at which gates an unverified claim becomes load-bearing.

Hallucination in hypothesis generation is close to free. You are paying for volume, and the wet lab is the filter. A fabricated binding interaction that fails an assay has cost a plate and a week.

Hallucination in a literature synthesis that justifies a target selection is expensive but recoverable. It costs a programme, and it is discoverable months later by someone competent reading backwards.

Hallucination in a safety rationale, a dose justification, or a regulatory submission is not recoverable. It ends careers, and before it does that, it harms patients.

So the human role cannot be to review everything, because that collapses the speed advantage that justified the model in the first place, and organisations that try it end up paying for the model twice. The human role is to own the gates where an error becomes irreversible, and to be accountable in a way that a system cannot be. Accountability is not a workflow property, it is a property of a person with a name and a licence.

Anthropic's reviewer agent concedes the point implicitly. A second model checking the first one's citations is an admission that the first one's output is not yet evidence. That is an honest piece of design, and it should be read as a statement about what the technology is, rather than as a solution to it.

The uncomfortable arithmetic is this. Verification cost scales with proximity to the patient. The entrant's advantage is largest at the distance furthest from the patient. Which is also where the least value sits.

What will actually decide it

Not model quality, and not compute, and not the size of the skills library.

The entrants who succeed will be the ones who solve for clinical and regulatory partnership early, most probably by licensing out at Phase I and never pretending to be a pharmaceutical company at all. That is an unglamorous business, and it is a viable one, and it is roughly what the biotechnology sector has done for thirty years with worse tools.

The ones who fail will be the ones who believed the discovery bottleneck was the bottleneck.