Anthropic released Claude Science beta on June 30, 2026. It is an AI workbench for scientists, designed to put literature analysis, scientific database queries, code execution, chart generation, compute management, and reproducibility into one environment.
It is not a normal chat window with a research theme. It is closer to an agent-powered research environment. Researchers can use it on local macOS or Linux machines, or connect through SSH or an HPC login node to existing lab compute resources.
What problem Claude Science tries to solve
Research workflows are often fragmented. A biomedical project may require PubMed searches, Jupyter notebooks, R scripts, PDB or UniProt data, cluster jobs, charts, and finally a paper or report.
Claude Science tries to connect those steps. It can:
- Analyze papers and preprints.
- Query specialist databases and research models.
- Execute multi-step research tasks.
- Generate charts, structure views, manuscripts, and analysis artifacts.
- Record the code, environment, and context used to produce results.
- Let reviewer agents check whether citations, calculations, and figures are traceable.
Anthropic emphasizes that every output includes an auditable history. That is crucial in research, where “looks right” is not enough: a result must be traceable back to data, code, and environment.
How it works
The entry point is a general coordinating Agent. It comes with more than 60 research-oriented skills and connectors covering genomics, single-cell biology, proteomics, structural biology, cheminformatics, and related areas.
It can also call other specialist Agents, or use specialist agents created by a research team. A reviewer agent checks references and calculations. When it finds numbers that cannot be traced, incorrect citations, or figures inconsistent with code, it flags them and attempts to fix the issue.
For research artifacts, Claude Science can render richer outputs, including:
- 3D protein structures
- genome browser tracks
- chemical structures
- charts and manuscripts
- the code and environment notes used to generate them
Researchers can ask in natural language to modify charts, such as removing grid lines or changing an axis to log scale. Claude Science edits the chart-generating code rather than only changing a static image.
Compute management is a key point
Many research analyses are hard not only because of code, but because of compute management. Protein folding, genomics pipelines, and large single-cell analyses may require GPUs, HPC clusters, or on-demand compute platforms.
Claude Science can draft compute plans, ask before accessing new resources, and let users review or withdraw decisions before submitting jobs. It can submit tasks to an existing lab HPC cluster or connect to Modal for on-demand compute.
Anthropic’s design goal is to keep data in its original system as much as possible. Claude Science can run on a lab laptop, Linux machine, or HPC login node. Large or sensitive datasets do not need to leave the original system wholesale; only the context needed for analysis is sent to Claude.
That matters for life sciences, medicine, and enterprise research, where proprietary data, unpublished data, and compliance requirements are common.
Preconfigured scientific capabilities
Claude Science is heavily preconfigured for life sciences. Anthropic mentions connections or support for UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, GEO, journals, preprint servers, and domain-specific open models.
It also uses skills from the NVIDIA BioNeMo Agent Toolkit, connecting to life-science models and libraries such as Evo 2, Boltz-2, and OpenFold3.
If a lab already has its own models, datasets, or pipelines, Claude Science can connect to them through connectors. Common pipelines can also be saved as reusable skills so later sessions inherit them automatically.
Early examples from Anthropic
Anthropic lists several beta use cases:
- Single-cell RNA sequencing analysis
- CRISPR screen design
- Protein structure prediction
- Cheminformatics
- Molecular epidemiology analysis
Manifold Bio used Claude Science to screen candidate targets for tissue-targeting medicines. Claude Science ranked candidates using criteria such as surface expression, trafficking, and safety, while incorporating context from Manifold’s internal historical projects.
At the Allen Institute, neuroscientist Jérôme Lecoq used it to build a multi-agent computational review template. About 20 custom skills read large bodies of papers, extracted core claims and quantitative results into an evidence state database, generated long review sections, and then used a reviewer agent to check accuracy and citation consistency.
At the UCSF Brain Tumor Center, Stephen Francis used Claude Science for glioma molecular epidemiology analysis. Anthropic says the team independently validated the results and estimated that the analysis time fell to about one-tenth of the original.
The common pattern is clear: Claude Science is better suited to long, data-heavy, reviewable scientific workflows than to simple Q&A.
Availability and application program
Claude Science is currently in beta, supports macOS and Linux, and is available to:
- Claude Pro
- Claude Max
- Claude Team
- Claude Enterprise
Team and Enterprise users need administrator enablement. Anthropic also offers a discounted Team plan for active scientific labs at academic institutions and nonprofit research organizations.
Anthropic will also support up to 50 Claude Science AI for Science projects, with up to $30,000 in credits for each project. Modal may provide up to $2,000 in compute resources for selected projects. Applications close on July 15, 2026; award notifications are planned for July 31, 2026; projects run from September 1 to December 1, 2026.
Important boundaries
Claude Science has a clear direction, but it is still beta. In research, AI output should not be treated as a conclusion, especially for citations, statistics, figures, and biomedical judgment. Researchers still need to verify results.
Its real value is connecting “find literature, write code, run jobs, draw charts, keep records, and review” into a traceable process. If artifacts, code, environment, and reviewer records are complete enough, AI is no longer only a faster writing or search tool; it can take part in a verifiable research pipeline.
In the short term, Claude Science is best suited to:
- Labs with large datasets and pipelines spread across many tools.
- Research teams that repeatedly generate charts, manuscripts, and reproducible analysis records.
- Life sciences, protein structure, genomics, cheminformatics, and other domains with complex data sources.
If Anthropic can stabilize reproducibility, permissions, auditing, and local/cluster execution, Claude Science may become one of the least chat-like products in the Claude line: more like a research console with memory, reviewers, and compute orchestration.
How a lab should try it
Claude Science is best tested with a real but bounded project. Do not move an entire lab workflow into it immediately. Start with a set of public papers, a reproducible dataset, an existing notebook, and an expected figure. Ask it to complete the loop: read literature, run analysis, generate a figure, and write the methods note.
Focus on three things. First, traceability: can the numbers in a figure be traced back to data and code? Second, environmental consistency: does the same task remain stable on another machine or when rerun? Third, citation quality: are references real and contextually correct, not merely formatted like a paper?
If the team has HPC or internal data sources, connect them in a second phase. Scientific data permissions are sensitive, so local paths, SSH permissions, database connections, and output directories should be tightly scoped before letting an Agent read or write.
Difference from ordinary Claude
Ordinary Claude is more like a research assistant: it can read papers, summarize, and explain code. Claude Science connects that assistant to a workbench: it can call specialist connectors, run code, generate artifacts, and preserve audit history.
So the question is not only “does it answer well?” The better question is “can it enter the research workflow?” Reproducibility, auditability, and integration with existing tools are the real differences between a scientific workbench and a chat window.
Original: Claude Science, an AI workbench for scientists, is now available