Anthropic has officially launched Claude Science, a specialized, standalone AI workbench app designed entirely to automate and accelerate scientific research.

Announced by CEO Dario Amodei alongside pharma executives from Novartis and Bristol-Myers Squibb at “The Briefing: AI for Science” event in San Francisco, the tool serves as an end-to-end digital lab assistant. It marks Anthropic’s first major domain-specific product vertical as the company scales up commercial revenues ahead of its highly anticipated $1 trillion IPO.

1. What is Claude Science? (App vs. Model)

It is crucial to note that Claude Science is a software application, not a new AI model.

The app runs on Anthropic’s existing model layer (such as Opus 4.8 and Sonnet 5). Instead of functioning as a basic chat window, the app gives Claude a dedicated desktop workspace equipped with over 60 pre-configured connectors, domain packages, and coding environments used natively by researchers.

2. Core Operational Capabilities

The platform is tailored primarily for advanced fields like genomics, single-cell biology, proteomics, structural biology, and cheminformatics:

  • Native Scientific Renderers: The interface displays 3D protein structures, chemical molecules, and genome browser tracks directly inside the workspace without requiring any external plugins or terminal installations.
  • Plain-Language Infrastructure Compute: Running complex analyses—such as folding a protein or crunching massive datasets—traditionally requires manually shifting data to computing clusters. Claude Science handles the engineering stack autonomously. It can write batch scripts, build environment containers, and manage computing pipelines over SSH via a user’s local laptop, an existing High-Performance Computing (HPC) node, or a Modal account.
  • Cross-Database Synthesis: Instead of forcing researchers to individually navigate fragmented data pools, a coordinating agent queries across 60+ major scientific databases simultaneously (including UniProt, PDB, Ensembl, ClinVar, and ChemBL) using tools like NVIDIA’s BioNeMo Agent Toolkit.
 [ Plain-Text Research Query ] ──► Specialist Agents query across 60+ Science DBs (PDB, UniProt)
                                                │
                                                ▼
 [ Claude Science Workbench  ] ──► Autonomously writes Python/R batch code to process data
                                                │
                                                ▼ (Compute Delegation)
 [ Infrastructure Execution  ] ──► Scales job across local CPU or remote HPC clusters via SSH
                                                │
                                                ▼
 [ Publication-Ready Output  ] ──► Renders interactive 3D structures + Fully Traceable Code Logs

3. The Reproducibility Mandate

Addressing a massive bottleneck in global research—where peer studies are frequently difficult to replicate—Anthropic built an absolute tracing net into the application:

Every artifact is auditable: When Claude Science generates a graph, data table, or draft manuscript chapter, the app bundles the exact underlying Python/R code, the execution environment configuration, and the chronological conversation history that built it. Any researcher on a team can open the file months later to perfectly replicate or defend the data.

Furthermore, scientists can use a plain-language annotator directly on a generated figure (e.g., “clean up this outlier axis”). The agent dynamically reads the backend code that produced the graphic and edits the file directly.

4. Availability & Enterprise Security Boundaries

The launch arrives on the heels of intense regulatory scrutiny regarding deep-tech AI. Because frontier models carry structural biological risks (such as bad actors attempting to engineer novel diseases), Anthropic is enforcing strict guardrails on its deployment:

ParameterApplication Specifications
Platform CompatibilityAvailable in public beta as a desktop app for macOS and Linux.
Tier AccessAccessible to users on paid individual and enterprise tiers (Pro, Max, Team, and Enterprise plans).
Admin ConstraintsTeam and Enterprise deployments require account administrators to explicitly enable the workbench access.
The Safety PipelineWhile running on standard models today, future integration with ultra-powerful frontier architectures (like Mythos 5) will be locked behind a rigorous, US-vetted Trusted Access Program specifically designed to isolate dangerous biological capabilities.

By acquiring biotech startup Coefficient Bio earlier in the year and actively partnering with global giants like Novo Nordisk to speed up pre-development molecule tracking, Anthropic’s new workbench moves AI out of the casual prompt box and straight onto the active laboratory bench.