Anthropic published early results from six drug-development partnerships on October 27, 2026, describing Claude’s involvement in protein engineering, antibody design, experimental optimization, regulatory paperwork and clinical-trial planning.
The strongest takeaway is the specificity of the work. Generate Biomedicines reported a shorter genetic component with measurable effects in laboratory tests. Divergence Labs generated roughly 35,000 antibody designs. Enveda reported cutting human document-review time by 89%.
Those numbers need separate scorecards. A design count measures output, receptor production measures a biological property, and review hours measure a workflow. Combining them into one claim about faster drug development would skip the steps connecting research activity to a useful medicine.
The results are reported by Anthropic and its partners, with no independent replication or audit in the reviewed sources. Most of the work is early or preclinical, while the paperwork and trial-planning examples concern development processes. None of these reported results demonstrates an approved medicine or improved patient outcomes.
Six Partnerships, Different Measures of Progress
Anthropic’s announcement names BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group. The available evidence supports detailed descriptions of five; The Column Group is identified as a partner, but the reviewed excerpts do not provide enough detail to characterize its project.
| Partner | Claude-assisted task | Reported result | What the measurement captures |
|---|---|---|---|
| Generate Biomedicines | Redesigning a genetic component for CAR T-cell therapy | One design about 70% shorter; roughly tenfold receptor production in laboratory tests | Design compactness and laboratory cell behavior |
| Divergence Labs | Generating antibody designs for a food-allergy approach | About 35,000 designs in roughly a month; early binding and IgE-removal hits | Computational output and experimental activity |
| BigHat Biosciences | Optimizing two antibody fragments | About 1,010 variants made and tested over ten weeks; improved binding reported | Experimental throughput and molecular properties |
| Enveda | Reviewing manufacturing documents and drafting regulatory material | 1,061 pages reviewed; 89% less human review time | Document-review performance and labor time |
| Bristol Myers Squibb | Assessing a clinical-trial plan and estimating success prospects | An initial analysis that the company said would otherwise take days | Planning assistance, without a quantified accuracy result |
| The Column Group | Partnership identified | No task-level result supported by the reviewed excerpts | Insufficient detail for comparison |
The design and laboratory figures come from Anthropic’s announcement and eWeek’s account. Becker’s Hospital Review supplies the document-review and trial-planning details.
My view is that this separation is the most useful way to read the announcement. The examples show several possible entry points for AI, but they cannot be ranked fairly by the size of their headline numbers. Thirty-five thousand proposed designs and an 89% reduction in review time answer different operational questions.
Generate Biomedicines: Shorter Instructions, More Receptors
Generate Biomedicines is working to extend CAR T-cell therapy into solid tumors. CAR T treatments use engineered immune cells with receptors that recognize cancer cells. Generate’s project focused on shortening part of the therapy’s genetic instructions, creating room for additional instructions suited to other cancers.
According to Anthropic, six Generate scientists used Claude Code to propose new designs. One was about 70% shorter than the standard version.
That percentage refers to the component’s length, not a reduction in development time. Likewise, the roughly tenfold figure refers to the number of receptors the immune cells produced, not the number of receptors researchers evaluated.
The reported laboratory results went beyond compactness: cells using the design produced roughly ten times as many receptors, killed tumor cells more efficiently and remained active longer.
For life-science teams, the useful pattern is a constrained engineering objective followed by several experimental readouts. The team wanted a smaller component, then checked whether that change preserved or improved relevant cellular behavior.
I would treat the combination of measurements as more informative than compactness alone. A shorter sequence would offer little value if the engineered cells lost function. Here, the reported functional results support further investigation, while leaving safety, performance in the body and clinical benefit as separate questions.
Divergence Labs: Design Volume Needs a Testing Denominator
Divergence Labs is developing a food-allergy approach targeting immunoglobulin E, or IgE. Some IgE circulates in blood; some is attached to immune cells. The company wants to detach the cell-bound IgE, which requires a design with two functional parts: one that binds IgE and another that helps remove it from the cell.
Claude translated that goal into a design plan and then generated roughly 35,000 antibody designs in about a month, largely autonomously, according to Anthropic’s account.
Early experiments found:
- Three designs that could bind IgE.
- Twenty designs that could dislodge IgE.
- Three of those twenty that appeared to dislodge more IgE than the standard medication used for comparison.
The next step described in the announcement is to combine the two functions into a drug design. Finding activity in each arm separately leaves an important integration task.
The key measurement gap is the testing denominator. eWeek notes that Anthropic did not disclose how many of the 35,000 designs were physically tested. Dividing the reported hits by the total design count would therefore give a misleading experimental success rate. The same report notes that the comparison medication was not named.
A reproducible workflow should track the entire selection process: designs generated, designs shortlisted, molecules successfully made, molecules tested and hits confirmed. That would show whether AI improved candidate selection or mainly expanded the pool of possibilities. Both could be useful, but their laboratory demands and economics differ.
BigHat Biosciences: Optimization Depends on the Lab Loop
BigHat Biosciences paired Claude with its automated laboratory workflow to improve two antibody fragments. One was designed to target cancer cells; the other was intended to help a drug remain in the body longer.
The optimization goals covered three properties:
- Binding: how tightly a fragment attaches to its target.
- Stability: whether it tolerates conditions involved in manufacturing and storage.
- Aggregation: whether molecules clump together.
Anthropic reports that BigHat scientists and an Anthropic researcher worked with Claude, connected to BigHat’s Reccy lab-data system. Claude proposed molecular changes, and the laboratory made and tested about 1,010 variants over roughly ten weeks. eWeek’s summary reports that the best variants bound their targets more than ten times as tightly.
The workflow detail matters: the model had a connection to experimental data, while the lab supplied physical measurements. This is a more concrete operating model than asking a chatbot to suggest a promising molecule without a defined testing process.
The 1,010-variant count describes throughput. Stronger binding describes one molecular property. Neither captures the full balance among binding, stability, aggregation and other requirements for a usable drug.
For teams considering a similar setup, I would prioritize traceability between each proposed change and its measured result. That record makes it possible to identify which changes helped, which introduced tradeoffs and whether later iterations are genuinely improving the candidates.
Enveda: A Clearer Workflow Measurement
Enveda’s example concerns manufacturing documentation, where work can be checked against existing records. Becker’s Hospital Review reports that manufacturing chemists had been spending up to 75% of their time on paperwork.
Claude reviewed 1,061 pages and flagged all 11 issues human reviewers had caught, plus seven additional issues. Human review time fell by a reported 89%, from more than 25 hours to under three.
This is the most directly operational result in the group: a defined body of documents, a comparison with human findings and a measured labor requirement. It gives another team a relatively clear starting point for designing its own evaluation.
Frequently Asked Questions
3 questions
1Did Claude Produce a Validated Medicine?
No. The reported results concern early research, laboratory measurements and development workflows. They do not demonstrate a newly approved medicine or improved patient outcomes.
