Table of Contents
01What Claude Science is
Anthropic released Claude Science as a beta app for scientists on June 30, 2026. As TechCrunch reported from the launch, Anthropic says it is not a new model. According to the Claude Science product page, it runs the Claude models already in a user's plan and adds scientific tools, database connections and compute integrations around them.
Anthropic treats it as a flagship. MIT Technology Review noted that the company placed it alongside Claude Code and Claude Cowork, and STAT reported that Dario Amodei expects it to change the life sciences the way Claude Code changed programming.
The app installs wherever the data lives: a laptop, a lab Linux box, an HPC login node or a cloud VM. Raw datasets stay on those machines, and only the context each step needs goes to Claude. Users talk to a coordinating agent that can start specialist agents for genomics, single-cell work, proteomics, structural biology and cheminformatics.
It can query more than 60 scientific databases, including UniProt, PDB, Ensembl, ClinVar, ChEMBL and GEO, and Anthropic's Claude Science product guide lists roughly 150 curated skills. Through NVIDIA's BioNeMo Agent Toolkit, it also reaches open biology models such as Evo 2, Boltz-2 and OpenFold3.
02How it speeds up research
The product guide names six kinds of work that fill a biologist's day: literature review, experiment planning, data management, data analysis, figure generation and scientific writing. Claude Science goes after the friction between them, such as database schemas, file-format conversions and package conflicts, and leaves the design decisions and the interpretation to the scientist.
01. Literature Review
Automated semantic sweeps across arXiv, PubMed, and BioRxiv, synthesizing core claims into structured evidence matrices.
02. Experiment Planning
Protocol drafting, control group formulation, reagent calculations, and compatibility checking across wet-lab steps.
03. Data Management
Schema mapping, FASTQ/BAM/VCF format conversions, barcode de-multiplexing, and automated pipeline hygiene.
04. Exploratory Data Analysis
Persistent kernel state in Python and R; interactive clustering, differential expression, and pathway enrichment without reloads.
05. Figure Generation
Publication-ready Matplotlib, Seaborn, and ggplot2 generation with automatic visual self-inspection and closed-loop outlier filtering.
06. Scientific Writing
Live Markdown and LaTeX drafting next to data artifacts, ensuring every number and table links directly to execution traces.
Several design choices cut out waiting and repetition. Datasets load once into a persistent Python or R kernel, so agents explore the data instead of reloading it. Every figure an agent makes is fed back into its own context, which lets it spot an outlier cluster in its own plot and filter it before moving on.
When a job needs more hardware, the app writes the batch script and submits it to the lab's SLURM cluster over SSH or to a Modal account for on-demand GPUs, scaling from one GPU to hundreds. The scientist approves each new resource before the agent touches it.
Speed only helps if the results can be checked. Each figure, table or notebook carries four layers of history: a plain-language description, the exact code, the conversation that led to it and a snapshot of every package version. Months later, anyone on the team can see how a number was produced by opening the artifact itself.
03Agentic loops and open questions
At the launch event, Zubair Jandali, Anthropic's head of healthcare and life sciences go-to-market, described science as a cycle of designing an experiment, running it, analyzing the data and asking the next question. He argued that the cycle stalls at the analysis step, the part that happens at a keyboard.
“The scientific method is a loop, too.” Zubair Jandali, Anthropic, quoted by R&D World at the Claude Science launch
The same report notes that Boris Cherny, who created Claude Code, now says his job is writing loops that keep an agent working until a set condition is met, rather than writing individual prompts. Claude Science builds that pattern into each session:
Plan
The agent drafts a step-by-step plan and waits for approval. The plan stays on screen as a checklist the scientist can edit or narrow while work runs.
Act
Agents run code, query databases and submit compute jobs, asking permission before reaching a new website, folder or machine.
Check
A separate reviewer agent reads the session as it runs and flags citations it cannot verify, numbers it cannot trace and figures that don't match their code. The working agent fixes them before it finishes.
Branch
The scientist can fork the session at any point to test two approaches side by side without losing the original thread.
A loop like this lets a scientist hand agents a question that has no answer yet and have them gather, test and check evidence step by step. At launch, Alexander Tarashansky, who led the product's development, showed Claude Science identifying new drug candidates for phenylketonuria, a rare genetic disease, without step-by-step human direction.
Anthropic still keeps people in the loop for decisions with real-world consequences. Eric Kauderer-Abrams, its head of life sciences, has said Claude should not run open-loop on tasks such as writing clinical trials or communicating with the FDA.
04Automating literature review and documentation
The clearest time savings so far come from the paperwork of science. Jérôme Lecoq, a neuroscientist at the Allen Institute, built a multi-agent review template from about 20 custom skills. Sub-agents read thousands of papers, pull out each paper's central claim and key quantitative finding, and store both in an evidence database.
The pipeline then plans the narrative, hands each section to its own sub-agent and builds cross-study figures straight from the database. Every writing agent is paired with a reviewer agent that checks accuracy and citation fidelity. A review like this used to take Lecoq's team up to two years. He now has about 10 of them, many longer than 100 pages.
Documentation happens as a side effect of the work. Because every artifact stores its code, environment and reasoning, the record of what was done exists before anyone sits down to write a methods section. Scientists draft manuscripts next to the analysis, with live Markdown and LaTeX previews.
Figure revisions take a sentence. A scientist can click on a plot and ask for a log-scale axis, no gridlines or a colorblind-safe palette, and the agent edits the code that drew the figure instead of redrawing it from scratch. Trygve Bakken at the Allen Institute says on the product page that the app turns his biological questions into well-cited first-pass literature reviews and genomics analyses, ready for his team to critique.
05Results labs have reported
The results below come from Claude Science users and Anthropic's own announcements. None of them appeared in my Claude 101 article, which covered the ART enzyme system, the nine-loop amplitude, the quantum circuit designs and the scorpion toxin binders.
A virus contaminant hiding in RNA-seq data
Stephen Francis's lab at UCSF ran its bulk RNA-seq data through Claude Science, and one of the first findings was a laboratory virus contaminant in the samples. On the Claude Science product page, Francis says: “We spun our wheels on this for the better part of a year.”
Glioma risk genetics in a tenth of the time
Francis, an epidemiologist at the UCSF Brain Tumor Center, studies how thousands of small-effect germline variants combine to shape a person's risk of glioma. Anthropic's launch post reports that Claude Science ran full germline workups across several methods in about one-tenth of the earlier time, and his group validated the results independently.
Targets for tissue-targeting medicines
Manifold Bio designs medicines that home to a specific organ or cell type. It used Claude Science to nominate the targets for its latest experiments. For each tissue and target, the app assessed surface expression, trafficking and safety, then ranked candidates against criteria Manifold had learned from its own proprietary data.
About ten reviews that once took up to two years each
Lecoq's multi-agent pipeline at the Allen Institute has produced about 10 reviews, many over 100 pages, with citations checked by reviewer agents. His team is now working with domain experts to improve the critic agents.
Drug candidates for phenylketonuria
The launch demo showed Claude Science identifying new drug candidates for phenylketonuria on its own. Anthropic also announced an internal preclinical drug discovery program for neglected diseases that traditional drugmakers tend to pass over. R&D World points out that the company has not said what it will do with promising candidates.
Novo Nordisk is testing it too. On September 16, 2026, Novo Nordisk announced a collaboration with Anthropic under which it will test Claude Science in specific research and development workflows.
06Where the limits are
Biology is harder to speed up than code. At launch, Amodei agreed with medicinal chemist Derek Lowe that biology is a messy evolved system, and said nobody knows for sure whether AI will become a general scientific collaborator, although he sees early signs. Lotte Bjerre Knudsen, Novo Nordisk's outgoing chief scientific advisor, added that the field will probably be listening to Lowe for another ten years.
Privacy has a boundary too. Raw datasets stay on the lab's own machines, but the content of prompts and model responses is processed by Anthropic under its standard retention terms. For biosecurity, the product guide says every agent carries biosecurity rules in its system prompt, a per-turn classifier runs that users cannot switch off, and the app went through external red-teaming before release.
07Getting access
Claude Science is in beta on macOS, Windows and Linux for Pro, Max, Team and Enterprise plans. On Team and Enterprise, an admin has to switch it on first.
For academic and nonprofit labs, Anthropic opened 10,000 seats on a Claude team plan for scientists on August 27, 2026, free or discounted for one year. The team plan for scientists includes Claude Science, switched on by default.
Anthropic is also funding up to 50 Claude Science projects through its AI for Science program with up to $30,000 in credits each, and Modal is adding up to $2,000 in compute for selected projects. Applications closed on July 15, and the projects run from September 1 to December 1, 2026.
08References
- Claude Science, an AI workbench for scientists
- Claude Science (beta) product page
- Claude Science product overview
- Claude Science is Anthropic's newest flagship product
- Anthropic releases Claude Science, a product aimed at researchers
- Claude Science bets on workflow, not a new model
- Anthropic says AI can run science experiments now
- NVIDIA launches BioNeMo Agent Toolkit
- Modal integration brings scalable compute to Claude Science
- Expanding our support for scientists
- Claude team plan for scientists
- Novo Nordisk taps Anthropic's Claude to speed drug discovery