- Anthropic launched Claude Science in public beta on June 30, 2026, directly integrating NVIDIA‘s BioNeMo Agent Toolkit to bring GPU-accelerated models including Evo 2, Boltz-2 and OpenFold3 into a single research environment.
- Scientists can use natural language prompts to orchestrate complex workflows across more than 60 scientific databases, removing the need for manual model configuration or specialist coding.
- Claude Science produces auditable output histories with preserved code and environments, addressing reproducibility requirements that matter most to enterprise research and pharma teams.
Anthropic‘s Claude Science, launched in public beta on June 30, 2026, embeds NVIDIA’s BioNeMo Agent Toolkit directly into a natural language research environment, letting scientists orchestrate GPU-accelerated workflows without writing a line of configuration code. The integration puts models like Evo 2, Boltz-2 and OpenFold3 within reach of a conversational prompt. For drug discovery and genomics teams, the practical implication is a shorter path from hypothesis to computational result.
Claude Science Launches with BioNeMo Integration
The platform pairs Anthropic’s language model with NVIDIA’s BioNeMo Agent Toolkit, which packages domain-specific computational models as callable programmatic skills. A researcher can describe an experimental goal in plain English; Claude Science interprets that intent, selects the appropriate tools, formats the required inputs, dispatches work across NVIDIA compute resources and returns finished output for review. Tasks that previously required manual pipeline construction and specialist coding become a directed conversation.
The integration surfaces NVIDIA’s accelerated libraries and NIM microservices inside the research environment, abstracting low-level infrastructure from the scientist. The stated goal is a faster, tighter loop between human scientific reasoning and machine-accelerated computation.
BioNeMo Agent Toolkit: Powering Scientific AI
The BioNeMo Agent Toolkit is a suite of domain-specific tools built for agentic life sciences workflows. Its packaged skills cover biomolecular modelling across distinct research phases: Evo 2 for genomic sequence work, Boltz-2 and OpenFold3 for protein structure prediction, and supporting cheminformatics libraries for small-molecule analysis. Each skill is optimised for GPU execution, so agents can gather evidence, reason across findings, run computational experiments and recommend next steps without leaving the environment.
According to NVIDIA, 18 of the top 20 global pharmaceutical companies already deploy BioNeMo in production, giving the toolkit an established footprint in enterprise drug discovery before Claude Science was announced.
Claude Science: An Agentic Workbench for Researchers
Claude Science connects those computational capabilities to more than 60 scientific databases, including UniProt, PDB and ChEMBL, through preconfigured agents that handle retrieval and synthesis. The workbench is available in beta on macOS and Linux, with support for local deployment, remote access via SSH or connection to high-performance computing login nodes. That deployment flexibility matters for teams handling sensitive data: datasets can remain within a lab’s own infrastructure rather than moving to an external service.
Reproducibility is a specific design target. The platform preserves code, execution environments and message history alongside every output, producing an auditable record of how a result was generated. For enterprise research teams and pharma organisations where regulatory validation of computational methods is non-negotiable, that audit trail is the feature most likely to determine adoption.
Accelerating Complex Life Science Workflows
The inhibitor design pipeline for oncology targets is the clearest demonstration of what the integration changes in practice. A scientist identifies a cancer-causing antigen mutation, then prompts Claude Science to design potential inhibitors. The platform calls BioNeMo’s relevant models and NVIDIA NIM microservices, runs inhibitor prediction and optimisation across GPU resources, and returns ranked candidates for human review. Steps that previously required separate toolchains, manual data formatting and specialist engineering support collapse into a single directed workflow.
Other use cases demonstrated during the beta phase include single-cell RNA sequencing analysis, CRISPR screen design and protein structure prediction. The common pattern across all of them is the same: an agent selects the right computational tool, handles the data pipeline and returns a finished result, with the researcher focused on the scientific question rather than the infrastructure behind it.
Implications for Drug Discovery and Beyond
For enterprise life sciences teams, the most consequential aspect of Claude Science is not the conversational interface but the audit infrastructure underneath it. Drug discovery organisations already running NVIDIA BioNeMo in production now have a path to layer natural language orchestration on top of existing compute investments without rebuilding pipelines. The preserved environments and output histories address a real compliance concern: demonstrating that an AI-assisted result is reproducible and traceable.
Anthropic says it is actively seeking feedback from the research community during the beta period to expand domain-specific agents and refine enterprise integrations. How widely the platform is adopted beyond early beta participants is harder to verify from public material at this stage. What is clear from the announced architecture is that the combination of auditable outputs, flexible deployment and pre-integrated GPU models removes several of the friction points that have slowed enterprise AI agent adoption in regulated research settings. For more analysis on enterprise AI strategy, visit our Enterprise AI section.
