Claude Science: Anthropic's Dedicated Workbench for Life Sciences Researchers

On June 30, 2026, Anthropic released the beta of Claude Science — a standalone desktop application for macOS and Linux aimed at researchers in the life sciences. This is a separate tool, not part of claude.ai; it functions as a specialized research environment that consolidates workflows previously scattered across dozens of tools.

Why Claude Science Is More Than Just Another AI Chat

Standard AI assistants provide responses as text; the researcher then downloads code, runs it in a separate environment, copies results back, and manually verifies whether citations are correct. Claude Science eliminates that cycle.

The application renders 3D protein structures, genomic tracks, and chemical structures directly within the interface — without exporting to PyMOL or IGV. Every generated plot comes with an auditable artifact that packages the exact code, library versions, and the full conversation history that led to it. Reproducibility stops being a good intention and becomes the default.

On top of that sits a Reviewer agent that independently checks mathematical calculations and citations before a researcher saves or shares findings. Sessions can be branched — forked — so a researcher can compare two approaches without losing context.

More Than 60 Skills for Specialized Domains

Claude Science ships with more than 60 curated skills and connectors grouped by research domain:

• Genomics — sequence analysis, variant calling, population genetics
• Single-cell analysis — cell clustering, pseudotemporal reconstruction
• Proteomics — quantification, modification identification
• Structural biology — prediction and comparison of three-dimensional structures
• Cheminformatics — molecular similarity search, bioactivity prediction

Natively available databases include UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, GEO, and PubMed, as well as preprint servers.

NVIDIA BioNeMo: Speed That Changes Practicality

A key architectural element is integration with the NVIDIA BioNeMo Agent Toolkit and its associated models:

• Evo 2 — genomic sequence modeling
• Boltz-2 — protein structure prediction
• OpenFold3 — third-generation structural analysis

NVIDIA’s companion post accompanying the announcement cites concrete speedups from the BioNeMo ecosystem: Parabricks reduces genomic pipelines from hours to minutes, RAPIDS-singlecell cuts single-cell analysis runtime from 52 minutes to 25 seconds, and nvMolKit accelerates certain molecular computations by up to 3,000×. According to NVIDIA, 18 of the 20 largest pharmaceutical companies already use BioNeMo infrastructure.

Compute Management Without DevOps Knowledge

Researchers who do not want to administer infrastructure can use Claude Science with four compute tiers:

1. Local laptop (macOS/Linux)
2. SSH access to a remote server
3. HPC cluster (login node)
4. On-demand GPU via the Modal platform — with up to $2,000 in introductory credits

This range means the same tool serves a PhD student on a MacBook and a group with access to institutional clusters, with no change to the interface or workflow.

Access and Costs

The beta is open to subscribers on Pro, Max, Team, and Enterprise plans. Academic and non-profit laboratories can apply for a discount on the Team plan. For selected projects from the AI for Science program, up to $30,000 in credit support is available; applications are accepted through July 15, 2026, with programs running from September 1 to December 1, 2026.

The most important architectural decision remains the one around privacy: because Claude Science supports execution on local or institutional infrastructure, sensitive clinical or genomic data never has to leave the researcher’s network. For institutions under regulations such as HIPAA or GDPR, this is not a minor detail — it is a prerequisite for even considering the tool.

Anthropic positions Claude Science as the answer to a concrete problem: life sciences researchers spend a significant share of their working time integrating tools rather than doing science itself. Whether a workbench with 60+ skills and reproducible artifacts will shift that ratio is what the beta period — which is just beginning — will reveal.