Anthropic Puts Claude Inside Drug Labs, Cutting Review Time by 89%
Anthropic details Claude's work across six biotech partnerships and its own in-house drug programs, with early wins in CAR-T, allergies, and regulatory paperwork.
- Anthropic announced deep drug development partnerships with six organizations plus its own in-house programs.
- Partners: BigHat, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines, The Column Group.
- Generate's Claude-designed CAR-T component was 70% shorter and 10x more effective at receptor expression.
- At Enveda, Claude cut manufacturing document review from 25 hours to under three, an 89% reduction.
- Anthropic acquired Mithril Medicine, a solo biotech targeting rare childhood muscular dystrophies.
- First human clinical trials from Anthropic's internal pipeline could begin as early as late 2028.
Anthropic is expanding Claude’s role in life sciences through collaborations with six organizations and its own drug-discovery programs. Anthropic’s announcement covers laboratory research, manufacturing records, regulatory drafting, clinical-trial planning, company formation, and treatments for rare childhood muscular dystrophies.
Drug development combines target selection, molecule design, laboratory testing, manufacturing, regulatory review, and clinical trials. Each stage produces specialized data and documents, often across separate teams and vendors. Anthropic’s strategy places Claude throughout that pipeline, including the administrative work that consumes much of a program’s time and budget.
Six partners, six bottlenecks
The collaborations span biotechnology startups, a pharmaceutical company, and an investment firm. Anthropic presents partner-selected case studies without a shared benchmark or independent replication.
| Partner | Claude’s role |
|---|---|
| Generate Biomedicines | Engineering part of a synthetic receptor for CAR-T cell therapy |
| Divergence Labs | Designing antibodies intended to treat food allergies |
| BigHat Biosciences | Optimizing drug candidates with an automated laboratory system |
| Enveda | Reviewing manufacturing records and drafting regulatory documents |
| The Column Group | Building a biotechnology company around AI-assisted operations |
| Bristol Myers Squibb | Assessing early clinical-trial plans and estimating their probability of success |
At Generate Biomedicines, six scientists used Claude Code to design part of a chimeric antigen receptor, the synthetic receptor that helps engineered immune cells recognize cancer. One design was about 70% shorter than the standard version. In laboratory tests, cells using it produced roughly 10 times as many receptors, killed tumor cells more efficiently, and remained active longer.
Divergence Labs used Claude to generate about 35,000 antibody designs in one month. The company identified three designs that bound to immunoglobulin E, or IgE, the antibody involved in allergic reactions. It also found 20 designs that dislodged IgE from immune cells, including three that reportedly outperformed the standard medication on that measure.
BigHat paired Claude with Reccy, its automated laboratory platform. The strongest candidates bound their targets more than 10 times as tightly, increased one cancer candidate’s temperature stability by 18°C, and slowed clumping in a long-acting drug candidate by more than 100-fold.
Paperwork produces the clearest numbers
Enveda’s manufacturing chemists spent up to 75% of their time preparing and reviewing documents, according to the company. Enveda connected Claude to its manufacturing records and used the model to identify inconsistencies in good manufacturing practice documentation, which records how a drug is produced and controlled.
In a test covering 1,061 pages, Claude found all 11 issues previously identified by human reviewers and flagged seven additional issues they had missed. Claude-assisted review took less than three hours, down from over 25 hours, an 89% reduction.
Claude also prepared first drafts of two regulatory documents. One draft took four hours to produce, compared with about three months of manual work. Enveda’s standard human review process found one minor error in the other draft.
A production deployment would need validated software, access controls, audit logs, traceable source material, and accountable expert approval. The reported tests measure review assistance and first-draft speed; they do not establish that Claude can independently approve manufacturing records or regulatory submissions.
Anthropic takes on drug discovery
Anthropic acquired Mithril Medicine, a one-person company developing treatments for two rare forms of muscular dystrophy. Founder Cody Tranbarger used Claude to coordinate contracts, logistics, and experiments across 11 external laboratories.
According to Anthropic, Mithril advanced a potential treatment from an initial idea to promising mouse results in less than 18 months. Tranbarger is continuing the research inside Anthropic and plans to begin several additional programs.
Mouse results remain an early step in development because many candidates later fail on safety, dosing, or efficacy. Anthropic says a candidate could enter human trials in late 2028 at the earliest if the preclinical work continues to succeed. A Phase 1 trial would primarily assess safety and dosing.
The partnerships feed the product
Anthropic describes the collaborations as a product-development loop: scientists expose workflow failures, Anthropic adjusts Claude and its tools, and those changes flow into products for other researchers. The company has released Claude Science, a research workbench, and a Life Sciences Verification Program that gives approved organizations controlled access to more permissive safeguards for legitimate biology work.
The Bristol Myers Squibb collaboration applies Claude to early clinical-trial plans and estimates of their probability of success. Anthropic provides no validation metrics for those estimates, leaving calibration, source traceability, and human oversight as central requirements for any operational use.
The Column Group is also creating Pleiades, a biotechnology company designed to run more drug programs with a small team. Claude will support scientific and administrative tasks, testing whether lower operating costs can make a broader portfolio of experimental treatments financially viable.
What developers can extract
For developers working in regulated or scientific fields, the Enveda and Mithril examples expose several recurring system requirements:
- Grounding and provenance: Outputs need links to controlled source documents, data versions, and experimental records.
- Bounded tool access: Agents coordinating laboratories or vendors need explicit permissions, spending limits, and approval gates.
- Workflow-level evaluations: Tests should measure missed issues, false alarms, review time, correction rates, and downstream failures.
- Expert review: Specialists remain accountable for manufacturing decisions, regulatory submissions, experiments, and clinical plans.
- Auditability: Regulated deployments need durable records of prompts, retrieved evidence, model outputs, edits, and approvals.
Anthropic does not disclose model versions, prompts, tool schemas, costs, full error distributions, or evaluation datasets for most of the case studies. Teams therefore cannot reproduce the reported results from the announcement alone.
Document-heavy workflows similar to Enveda’s also appear in legal services, finance, compliance, and engineering. The transferable pattern combines private data retrieval, structured review, measurable evaluations, and expert sign-off.
Human trials remain years away
Nearly all the reported drug work remains preclinical, and none of the candidates described has been tested in people. Human safety data, controlled efficacy results, manufacturing validation, and regulatory review will determine whether the early laboratory gains translate into viable treatments.
Comments
