Finn Brooks: Clara, quick thing before we start — did you see the actual cost on the zoonotic spillover experiment they're citing?
Clara Bennett: Twenty-six dollars. Yes.
Finn Brooks: Twenty-six dollars! And I've been trying to figure out if that number is supposed to be impressive or slightly terrifying, because — okay, so today we are getting into Claude Science, Anthropic's new scientific workbench, launched June 30th in beta, and there is a lot happening here. The headline experiment: one researcher, 490 papers on zoonotic disease spillover, twenty-six dollars, and they surface 864 missing relationships out of 915 that official scientific ontologies just didn't have. That's the pitch.
Clara Bennett: And the bigger pitch is Zubair Jandali saying Claude doesn't help with the work — it runs it.
Finn Brooks: Which — no but that phrase, 'run the work, not help with it,' that's what I keep getting stuck on, because that's not a feature description, that's like... a replacement claim. And Anthropic also just announced they're starting their own preclinical drug discovery programs. So they're the tool vendor and now they're also the — wait, are they a pharma company?
Clara Bennett: Not a pharma company. That's the reframe I want to land. Think of it like a contractor who shows up with every tool already organized in one van. The tools aren't new — you've seen a drill before. But you'd never have rented all of them yourself. The van is Claude Science. The contractor is Claude Opus 4.8. Same model that was already there.
Finn Brooks: Wait — Anthropic actually said that? Like, confirmed it's just Opus 4.8 underneath?
Clara Bennett: Explicitly. No special model access, no capability upgrade — their words. It's a workflow harness. Now, that's the pitch, but it's also, mm, immediately the vulnerability, because if the power is sixty-plus pre-wired databases — UniProt, PDB, ChEMBL — what actually stops OpenAI from wiring those same databases into their interface?
Finn Brooks: Okay that's — yeah, that's the thing I keep bouncing off. Like, is the moat the plumbing or the... wait, is there even a moat?
Clara Bennett: Claude Code is the model they're betting on here. That product succeeded not because the underlying model was better than GPT — it was workflow integration that drove adoption. So the argument is: first-mover on the assembled van beats raw capability. The honest question is whether that holds in science the way it held in coding.
Finn Brooks: And those 864 relationships from the zoonotic experiment — those still need a scientist to actually confirm they mean something, right? The van delivered them but someone still has to check if the drill holes are in the right wall.
Clara Bennett: That's exactly right — and that's actually the tell I want to get to. Eric Kauderer-Abrams, Anthropic's head of life sciences, said the drug programs exist so Anthropic can gain firsthand experience using Claude Science on real scientific problems.
Finn Brooks: Wait — he said that out loud?
Clara Bennett: Those are essentially his words. And I mean — sit with that for a second. If Claude Science were unambiguously production-ready, you don't need to run your own drug programs to validate it. That framing is, mm, accidentally an admission.
Finn Brooks: Oh that's — no, that's actually a devastating read. Like, the confidence of 'we run the work' and then behind the curtain it's 'we're still learning what that means in practice.'
Clara Bennett: Now here's where the financial scale makes it genuinely uncomfortable. Anthropic's annualized sales are approximately $42 billion — that's GSK territory. Valuation sits at $965 billion, which clears every healthcare company except Eli Lilly. They're not a startup learning by doing. They're a peer-scale actor who now has — wait, actually, the phrase to use is structural access. They see what every pharma partner is running through Claude Science before those partners see their own results.
Finn Brooks: Okay that's the part that — dude, Basecamp Research is running antibiotic design through Claude Science. Vaccine prediction models. Targeting drug-resistant infections tied to nearly five million deaths a year. And STAT News broke that Anthropic is simultaneously building its own drug candidates.
Clara Bennett: The 'neglected diseases' framing is doing a lot of work there. It's supposed to signal non-competition with large pharma customers. But the other read — the uncomfortable one — is that Anthropic is cherry-picking targets where platform advantage is most decisive.
Finn Brooks: So the 'we're just learning' framing from Kauderer-Abrams is almost harder to sustain at $42 billion in sales than it would be if they were actually a scrappy startup.
Finn Brooks: Okay, I'll half-concede on 'run the work' — that might be overselling it. But '$26 to rewrite the ontology of a disease field' is still, like... I think that's the most interesting sentence in science this year. I'll stand there.
Clara Bennett: That's fair. And importantly — Claude for Life Sciences launched in October 2025, then Claude Science built on top of it this June. That's not a bet Anthropic is making. That's a bet they've been making for a while now. The real test isn't whether the workflow holds up. It's whether pharma still trusts them once they start winning at drugs.
Finn Brooks: Yeah. That's — actually kind of an uneasy place to land.
Clara Bennett: It is. Good thinking-through, though. Genuinely.
