Anthropic starts making its own drugs
-
Nº L
- Date
- 06 Jul 2026
- Issue
- 50
- Stories
- Six
- Editor
- ARC
Monday reset: Anthropic crossed from selling AI to running its own pipeline, and the benchmarks are starting to catch up to what agents actually do.
Anthropic enters drug development
Anthropic is developing drugs of its own, announced alongside Claude Science — a workbench wiring Claude into 60+ scientific tools and databases and capable of running full experimental loops. The move lands days after Claude Sonnet 5 shipped, and pushes Anthropic past the tool-vendor line into being a discovery operator. It's the clearest signal yet that frontier labs see more margin in owning pipeline assets than in licensing the models that generate them.
Claude Sonnet 5 lands
Claude Sonnet 5 shipped as Anthropic's most agentic model yet — planning multi-step tasks, calling tools, and running autonomously at near-Opus 4.8 quality for a fraction of the cost. The release is the substrate Claude Science and the new in-house drug program run on, and resets the price-performance ceiling agent platforms have been quoting all quarter.
Active learning for antibiotic screens
Active learning boosts generalizability in whole-cell antibiotic discovery, a bioRxiv preprint reports — models trained with iterative query selection find hits across bacterial species their training sets never covered. Narrows the gap between in-silico antibiotic hunts and the phenotypic screens that still dominate the field, where cross-species transfer has been the standing bottleneck.
Foundation models learn perturbations
Task-adapted foundation models recover perturbation-centric representations from single-cell data, capturing how cells respond to drugs and genetic edits rather than just what they look like at baseline. Shifts the useful axis of cell-atlas models from description to intervention.
PACE benchmarks agent capability
PACE proposes a proxy for agentic capability evaluation — a lighter-weight harness for scoring agents without running full end-to-end task suites. Cuts the cost of comparing agent releases, which matters as biology-facing agents proliferate faster than anyone can benchmark them properly.
Bounded-memory testbed for long agents
AgenticSTS stresses long-horizon LLM agents under bounded memory, exposing how they degrade past 20+ tool calls. Anchors a concrete measurement of the context-drift problem that long-running biology agents — literature triage, multi-day docking loops — hit in production.
Reply with your discoveries. A human reads them. Forward freely.
|