Anthropic makes AI-for-science a flagship product — and starts hunting its own drugs
Claude Science launches as a research workbench orchestrating 60+ scientific databases and tools, while Anthropic discloses its own internal drug-discovery pipeline for neglected diseases. Plus: Takeda's ~$600M bet on Insilico, an economist's case that AI drug discovery solved the cheap problem, and The Onion Desk.
Anthropic launched Claude Science on June 30 as a standalone research workbench — joining Claude Code and Claude Cowork as a flagship product — and simultaneously disclosed it is running internal drug-discovery programs for neglected diseases. The workbench itself isn't a new biology model; it coordinates specialist sub-agents across 60+ curated databases and tools, from UniProt and PDB to Jupyter notebooks and cluster terminals, with a dedicated reviewer agent checking citations and calculations before publication. The more interesting move is buried in the announcement: Anthropic courting pharma customers with one hand while competing against them in drug discovery with the other.
Lead — Claude Science and the neglected-disease pipeline
Product launchAnthropic launched Claude Science on June 30 — a top-line product in beta across subscription tiers, alongside Claude Code and Claude Cowork — and disclosed it is running internal drug-discovery programs targeting commercially unattractive, neglected diseases. Claude Science isn't a new biology-specific model; it coordinates specialist sub-agents across 60+ curated databases and tools, including UniProt, PDB, Ensembl, ChEMBL, GEO, PubMed, Jupyter, and cluster terminals, rendering 3D structures and genome-browser tracks while keeping an auditable output history. Beta partners report substantial time compression: Allen Institute researchers produced '100+ page literature reviews that previously took as long as two years,' and UCSF's Brain Tumor Center compressed a comprehensive germline glioma analysis 'to roughly a tenth of its prior time, with independently validated results.' For research leaders, Claude Science reframes build-versus-buy around operating-layer integration rather than raw model superiority — and the sharper story is strategic: Anthropic is simultaneously courting pharma customers and competing against them in neglected-disease drug discovery, a credibility play with latent conflicts.
Takeda writes a ~$600M check for Insilico's discovery engine
DealsTakeda entered a strategic collaboration worth roughly $600M for exclusive worldwide rights to drug candidates from Insilico Medicine's Pharma.AI platform — PandaOmics for target ID, Chemistry42 for generative small-molecule design, inClinico for trial-success prediction. It's another major-pharma bet on AI-native platforms over one-off asset acquisitions, with Insilico's credibility resting on rentosertib's positive Phase IIa in IPF.
An economist argues AI bet on the wrong bottleneck
OpinionSanta Clara economist Michael Santoro argues the AI-discovery decade compressed the wrong bottleneck. A BCG analysis of ~24 AI-discovered molecules shows Phase 1 safety success jumping to 80–90% versus a historical ~50%, but Phase 2 efficacy 'falls right back to the industry's usual ~40%.' With 15–20 AI-discovered programs reaching Phase III this year, his point is that the real opportunity is applying AI to efficacy, not building yet more molecule-design platforms.
"Good to great, not nothing to great": a builder's reality check
OpinionBioptimus CEO Jean-Philippe Vert offers a builder's reality check in the same vein: AI's proven advantage so far is time compression — target-to-candidate work down from 5–7 years to under 2, especially in oncology — while interpretability and data silos remain real limits. His secondary worry is that cheaper candidate generation floods already patient-constrained clinical pipelines, pushing the bottleneck downstream toward development and efficacy.
The quiet tell in Claude Science: it's built on someone else's engines
InfrastructureThe quiet tell inside Claude Science is that it calls NVIDIA's BioNeMo Agent Toolkit models — Evo 2, Boltz-2, OpenFold3 — as tools rather than shipping proprietary alternatives. That's a bet on orchestration value over foundation-model ownership, at exactly the layer Google DeepMind controls end-to-end. Structural advantage accrues to whoever owns both the orchestration and the underlying engines.
BridgeBio raises ~$1B in preferred equity
FundingBridgeBio raised roughly $1B in preferred equity from Sixth Street and KKR's HealthCare Royalty to fund three genetic-disease launches — a marker of where late-stage capital and structured, non-dilutive financing are flowing.
MHRA approves Wegovy for UK MASH
RegulatoryThe MHRA approved Wegovy for UK MASH treatment with moderate-to-advanced fibrosis, setting the GLP-1 label-expansion benchmark that AI-discovered metabolic programs will be measured against.
A former FDA regulator says industry reads the guidance too conservatively
RegulatoryA former FDA AI regulator suggests industry is interpreting existing guidance overly conservatively — implying the constraint on AI-evidence submissions may be interpretation rather than the rules themselves.
Edison Scientific and Population Health Partners build AI-agent-native biotechs
DealsEdison Scientific and Population Health Partners are collaborating to build AI-agent-native biotechs, where the strategic bet is company format itself rather than any single molecule.
The Onion Desk
Regulators unveil ten AI principles alongside a promised eleventh — due by 2029 — to clarify the original ten, birthing what one wag called self-referential compliance architecture. Meanwhile pharma executives express disappointment that AI compressed molecule design but still can't make drugs work in humans, with one team reporting their AI assistant refuses spreadsheets containing the word 'Ebola,' forcing euphemistic pandemic research. Full satirical dispatch on Substack.