On 30 June, Anthropic released Claude Science, an AI workbench for researchers. The capabilities are real and, in places, genuinely striking. It pulls sixty-odd scientific databases, coding tools and compute into a single environment; it renders 3D protein structures and genome tracks natively; it runs multi-step analyses across a lab’s own infrastructure and manages the cluster jobs itself. In beta, a neuroscientist at the Allen Institute used it to build review pipelines that turned a task once measured in years into one measured in weeks; a UCSF group ran a genomics workup in roughly a tenth of the time. Anthropic paired the launch with something bigger still: its own internal drug-discovery programme.
It is an impressive piece of engineering. It is also aimed squarely at the part of the problem that was never the problem.
This is a discovery tool. And discovery — the identification and design of the molecule — is the one gate the pharmaceutical industry already wins. As this publication argues in almost every issue, approval is not access, and the molecule was never the reason patients wait. AI is about to make the well-lit, well-funded, well-rehearsed part of the pipeline faster still, and leave the two dark gates exactly where they are.
The gate that was never the constraint
Recall where medicines actually go missing. The average wait between EU approval and a patient in Europe is 578 days. Fewer than half — around 43% — of the innovative medicines the EMA approves ever become available across the bloc. New launches in Europe have fallen 35% in the ten months since the United States tied its prices to the world’s lowest, as companies delay and withdraw to protect the American margin. None of that is a science problem. Not one day of that 578 is spent waiting for a protein to fold.
A tool that folds proteins in minutes does not shorten a pricing negotiation by an hour. A reviewer agent that checks citations does not persuade a payer to fund a therapy it has decided it cannot afford. The constraint on medicines reaching patients has, for years, lived entirely downstream of the laboratory — in reimbursement, in health technology assessment, in the reference-pricing cascade, in the launch-sequencing calculus that now keeps drugs out of low-price markets on purpose. AI, for now, is pointed at none of it.
So consider what happens if it works as advertised. If discovery gets ten times cheaper and faster, the funnel fills with ten times as many candidates — and they arrive at the same wall. More molecules, the same access machine. The gap between approved and available does not close. It widens, because you have accelerated the one stage that was never the bottleneck and touched nothing about the stages that are.
Faster discovery does not mean faster access. It means more approvals queued behind the same wait.
The part that actually matters for our world
There is a genuinely relevant feature here, and it is not the protein rendering. It is the audit trail.
Claude Science attaches to every figure the exact code, the computing environment, a plain-language account of how it was made, and the full message history — a reviewer agent flags untraceable numbers and figures that do not match their code. In regulated pharmaceuticals, this is the part that could matter, because adoption in our world has never depended on whether a tool works. It depends on whether its output can be validated, reproduced, and audited. The provenance chain — not the speed — is what determines whether AI-generated analysis is ever admissible in a regulatory dossier or an HTA submission. That is the development worth watching: not that Claude can draw a molecule, but that it can show its work in a form a regulator might one day accept. The bottleneck for AI in pharma was always going to be evidentiary, not computational — and this is the first serious attempt to build for it.
The tell is the neglected-disease programme
Which brings us to the most revealing part of the announcement, and the one that proves the point.
Alongside the workbench, Anthropic said it would run its own drug-discovery programmes targeting “neglected” diseases — conditions, in its life-sciences head’s own framing, outside the scope of what traditional pharma and biotech consider attractive targets, but carrying real burden. It is a serious and, on its face, admirable move.
But ask the question the phrase quietly begs. Why are those diseases neglected? Not because the science was impossible. Because no one will pay for the cure. These are the diseases we wrote about when we asked why the therapies that cure are the ones Europe can’t pay for — the sickle cell of Nigeria and the Democratic Republic of the Congo, the conditions with enormous human burden and no viable commercial price. They were never neglected for lack of a molecule. They were neglected for lack of a payer.
AI can now discover treatments for them faster than ever. And the single reason they were neglected in the first place — that there is no one at the end of the pipeline willing or able to pay — is the exact thing no model touches. You can automate the search for the compound. You cannot automate a willing payer, a reimbursement decision, or a health system that can absorb the cost.
AI has made the molecule cheaper to invent. It has done nothing to the reason the molecule so often never arrives. The cure was never the hard part — and the most advanced tools we have just made the easy part easier.
You can accelerate discovery to the speed of light. Until someone at the other end agrees to pay, “discovered” will keep meaning exactly as little as “approved” already does.
Beyond Approval publishes weekly strategic intelligence on market access, HTA, pricing, and regulatory shifts that reshape how medicines reach patients.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.