Claude Science: Anthropic's AI Workbench for Drug Discovery and Biomedical Research

Executive Summary

On June 30, 2026, Anthropic launched Claude Science — not a new model, but an entirely new category of product: an AI workbench designed specifically for scientific research. Released alongside Claude Sonnet 5 and the restoration of Fable 5, Claude Science represents Anthropic's most significant expansion beyond conversational AI, targeting the $3.8 trillion global scientific R&D market.

Claude Science integrates over 60 curated skills and connectors for genomics, single-cell analysis, proteomics, structural biology, and cheminformatics into a single research environment. It natively renders 3D protein structures, genome browser tracks, and chemical structures; manages compute from a laptop to hundreds of GPUs via Modal or HPC clusters; and employs a multi-agent architecture with dedicated reviewer agents that check citations, calculations, and figure-code consistency in real time. Every output carries an auditable history of how it was produced.

Early beta results are striking: a neuroscientist at the Allen Institute compressed a 2-year literature review process to weeks, producing 10 reviews exceeding 100 pages each with verified citations. An epidemiologist at UCSF achieved 10× speedup on germline variant analysis for glioma research. Manifold Bio used Claude Science to nominate drug targets end-to-end, assessing surface expression, trafficking, and safety across millions of candidate binders. Anthropic is also launching its own internal pre-clinical drug discovery program targeting neglected diseases, with up to $30,000 in credits for 50 external research projects (applications close July 15).

---

1. The Problem: Fragmented Scientific Toolchains

1.1 The Current State of Scientific Research

Scientific research — particularly in life sciences — is characterized by extreme tool fragmentation. A typical researcher's workflow spans:

• Dozens of databases (UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, GEO), each with unique schemas and query languages
• Multiple computing environments (local machines, HPC clusters, cloud VMs), each requiring separate provisioning and authentication
• Diverse software tools (Jupyter, R, terminal, domain-specific pipelines), each with its own environment management
• Manual reproducibility tracking — scientists must independently document how each figure, table, and result was produced

This fragmentation creates massive overhead. As Anthropic stated in their launch announcement:

"Scientific research is often tedious. Researchers must work across dozens of databases, each with their own schema, contend with file formats that require bespoke data pipelines and viewers, and transition between a roster of tools: PubMed, Jupyter, R, a cluster terminal, and more."

1.2 The Agentic Opportunity

The emergence of capable agentic models — documented extensively in Claude Sonnet 5 Agentic Mid Tier Model 2026 07 03 and Openai Gpt 56 Sol Terra Luna Subagent Ultra Mode Cyber Safeguards 2026 07 06 — creates an opportunity to unify these fragmented workflows. A sufficiently capable agent can navigate multiple databases, execute code across environments, and maintain context across multi-step research pipelines — effectively acting as a PhD-level research assistant with supercomputer access.

---

2. Architecture: The Multi-Agent Workbench

2.1 Core Components

Claude Science is built on a multi-agent architecture that extends beyond the single-agent chat paradigm:

Component	Role	Key Capability
Coordinating Agent	Generalist orchestrator	Parses plain-language queries, delegates to specialist agents, manages workflow
Specialist Agents	Domain-specific workers	Genomics, proteomics, cheminformatics, structural biology, single-cell analysis
Reviewer Agent	Quality assurance	Checks citations, validates calculations, verifies figure-code consistency, flags errors
User-Created Agents	Custom pipelines	Researchers can create and save their own specialist agents as reusable skills

2.2 The Reviewer Agent: Actor-Critic Architecture

One of Claude Science's most innovative features is the built-in reviewer agent — an implementation of the actor-critic pattern documented in reinforcement learning literature, applied to scientific workflow validation:

The reviewer agent specifically checks for:

• Incorrect citations — verifying that referenced papers actually support the claims made
• Untraceable numbers — ensuring every quantitative result can be traced back to its source code
• Figure-code mismatches — confirming that rendered figures match the code that produced them
• Self-correction — when errors are found, the system loops back to the specialist agent for correction

This architecture directly addresses one of the most persistent problems in scientific research: reproducibility. As demonstrated by Jérôme Lecoq at the Allen Institute, the actor-critic pattern enabled his team to produce 10 reviews exceeding 100 pages each, with all citations verified by reviewer agents — work that previously took two years.

2.3 Session Management

Claude Science maintains running sessions that hold context in memory, enabling:

• Single data load — Even massive datasets only need to be loaded once into the session
• Forkable workflows — Users can fork a session at any point to compare two approaches without losing the original thread
• Persistent context — Agents work within a continuous session rather than stateless API calls

---

3. Native Scientific Rendering

3.1 Rich Artifact Generation

Unlike standard AI chat interfaces that output text and code blocks, Claude Science natively renders scientific artifacts:

Artifact Type	Rendering	Use Case
3D Protein Structures	Interactive 3D viewer	Structural biology, drug binding analysis
Genome Browser Tracks	Scrollable genomic visualization	Variant analysis, regulatory element identification
Chemical Structures	2D/3D molecular rendering	Cheminformatics, drug design
Single-Cell Plots	UMAP/t-SNE embeddings	Cell type annotation, differential expression
Publication Figures	Editable, publication-ready	Manuscript preparation

3.2 In-Line Figure Editing

Users can request figure modifications in plain language:

"Remove the gridlines" "Change the Y-axis to log scale" "Highlight the top 5 differentially expressed genes"

The agent edits its own code and re-renders the figure, maintaining the full audit trail of what changed and why.

3.3 Reproducibility by Design

Every generated figure includes:

1. The exact code that produced it
2. The environment (dependencies, versions) used to run it
3. A plain-language description of how it was created
4. The full message history leading to its creation

This ensures that even months later, a researcher can understand the inputs, validate the methodology, and reproduce the result.

---

4. Compute Management: From Laptop to Supercluster

4.1 Flexible Infrastructure

Claude Science handles compute provisioning transparently, supporting three infrastructure tiers:

Tier	Infrastructure	Use Case
Local	macOS or Linux machine	Early-stage analysis, small datasets
HPC Cluster	Lab's existing cluster via SSH	Large-scale genomics, established pipelines
Cloud (Modal)	On-demand GPUs via Modal integration	Bursty workloads, GPU-intensive tasks

4.2 The Modal Integration

Anthropic partnered with Modal to provide elastic, on-demand compute for Claude Science:

• Automatic scaling — Workloads that outgrow local resources automatically route to Modal sandboxes
• Heterogeneous compute — Different pipeline steps can use different hardware (CPU for alignment, GPU for structure prediction)
• Shared storage — Modal Volumes eliminate expensive data movement between pipeline steps
• Reproducible environments — Modal Images ensure consistent dependencies across runs

As Modal described in their integration announcement:

"A virtual screen run against a large compound library that would take hours sequentially fans out across hundreds of Modal containers and comes back in minutes."

4.3 Compute Workflow

4.4 Data Privacy

Critical for sensitive research: data never leaves the researcher's infrastructure. Claude Science runs on the lab's own systems (laptop, Linux box, or HPC login node). Only the context needed for each analysis step is sent to Claude — not the raw datasets.

---

5. Pre-Configured Domain Knowledge

5.1 Scientific Databases

Claude Science connects natively to over 60 scientific databases, including:

Domain	Key Databases
Proteomics	UniProt, PDB, Reactome
Genomics	Ensembl, ClinVar, GEO
Cheminformatics	ChEMBL
Literature	PubMed, preprint servers
Domain Models	NVIDIA BioNeMo models (Evo 2, Boltz-2, OpenFold3)

5.2 NVIDIA BioNeMo Integration

Claude Science uses skills from NVIDIA's BioNeMo Agent Toolkit to connect natively to life sciences models and libraries:

• Evo 2 — Protein language model for evolutionary analysis
• Boltz-2 — Protein structure prediction
• OpenFold3 — Open-source protein folding (University of Washington IPD collaboration, 2× faster than prior generation)

As Jensen Huang stated at the BioNeMo announcement:

"Frontier models are the brains. BioNeMo is the scientific toolbox. Together, they give AI agents the skills of a PhD research assistant and the speed of a supercomputer."

5.3 Custom Skills and Connectors

Researchers can:

• Save any pipeline as a reusable skill
• Connect to their lab's preferred tools via connectors
• Have future sessions automatically inherit custom configurations

---

6. Real-World Beta Results

6.1 Case Study: Manifold Bio (Drug Target Nomination)

Company: Manifold Bio — designs tissue-targeting medicines that home to specific organs/cell types.

Workflow: For each tissue and target, Claude Science assessed surface expression, trafficking, and safety, ranking candidates against Manifold's internal proprietary data criteria.

Key finding: What set Claude Science apart from a general coding assistant was its ability to execute the entire pipeline end-to-end — gathering the right data, applying the right judgment, and incorporating context from past programs.

6.2 Case Study: Allen Institute (Literature Review)

Researcher: Jérôme Lecoq, neuroscientist.

Workflow: Built a multi-agent "computational review template" with ~20 custom skills:

1. Sub-agents read thousands of papers, extracting central claims and quantitative findings
2. Evidence stored in a structured database
3. Pipeline constructs narrative arc, writing section by section
4. Dedicated agents generate quantitative cross-study figures from evidence database
5. Actor-critic pairs ensure accuracy: one agent creates content, reviewer agent evaluates

Results:

• Before: 2 years to write such a review
• After: ~10 reviews, many exceeding 100 pages, with verified citations
• Current: Working with domain experts to refine AI-based critic agents

6.3 Case Study: UCSF Brain Tumor Center (Genomic Epidemiology)

Researcher: Stephen Francis, associate professor and epidemiologist.

Workflow: Molecular epidemiology of glioma — investigating how thousands of small-effect germline variants combine to shape individual susceptibility.

Results:

• 10× speedup — comprehensive germline workups across multiple approaches in roughly one-tenth the time
• Independently validated — the group confirmed Claude Science produced both rapid and robust analyses

6.4 Performance Comparison

Task	Traditional Timeline	Claude Science Timeline	Speedup
Major literature review	2 years	Weeks	~50×
Germline variant analysis	Baseline	1/10× time	10×
Virtual screening (large library)	Hours (sequential)	Minutes (parallel)	~60×
CRISPR guide design (genome-wide)	Days	Minutes	~100×

---

7. Anthropic's Internal Drug Discovery Program

7.1 A New Direction for AI Companies

In a move that signals Anthropic's long-term commitment to life sciences, the company announced it will deploy Claude Science in its own pre-clinical drug development programs, focused on neglected diseases — conditions that fall outside the commercial calculation of traditional pharmaceutical investment.

This is significant because:

• It moves Anthropic from tool provider to active researcher
• It targets neglected diseases (rare diseases, conditions with no commercial incentive), aligning with Anthropic's stated values
• It creates a flywheel: internal drug discovery programs generate real-world validation data that improves Claude Science

7.2 The AI for Science Research Program

Anthropic is funding external research through a structured program:

Detail	Specification
Projects funded	Up to 50
Credits per project	Up to $30,000
Modal compute	Up to $2,000 per project (total $100,000 pool)
Application deadline	July 15, 2026
Award notification	July 31, 2026
Project period	September 1 – December 1, 2026
Focus areas	Biology and biomedical research, cross-domain exploration

---

8. Availability and Access

8.1 Current Access

Plan	Access	Notes
Pro	✅ Beta	Direct access
Max	✅ Beta	Direct access
Team	✅ Beta	Admin must enable
Enterprise	✅ Beta	Admin must enable
Free	❌	Not available

8.2 Platform Support

• macOS — Native app
• Linux — Native app
• Remote access — Via SSH or HPC login node
• Web — Available at claude.com/science

8.3 Academic Pricing

Anthropic introduced a Team plan offering discounted seats for active scientific labs at academic institutions and nonprofit research organizations.

8.4 Community

Researchers can join the AI for Science Discourse community for updates, feedback, and collaboration.

---

9. Comparison: Claude Science vs. Alternative Approaches

9.1 Against General-Purpose Coding Assistants

Capability	Claude Science	Claude Code / Codex	Jupyter + LLM
Native 3D molecule rendering	✅	❌	❌
Pre-configured 60+ bio databases	✅	❌	Manual
Built-in reviewer agent	✅	❌	❌
Multi-agent orchestration	✅	Limited	❌
Compute auto-scaling (Modal/HPC)	✅	❌	Manual
Auditable figure history	✅	❌	Manual
Domain-specific skills	✅	❌	❌

9.2 Against Specialized Bioinformatics Platforms

Capability	Claude Science	Benchling	DNAnexus
Natural language interface	✅	❌	❌
Multi-agent reasoning	✅	❌	❌
Flexible compute (local/HPC/cloud)	✅	Cloud-only	Cloud-only
Literature integration	✅	Limited	Limited
Custom agent creation	✅	❌	❌
Cost	Included in subscription	$$$	$$$

---

10. Key Takeaways

1. New category, not just a feature: Claude Science is not a chatbot with science prompts — it's a purpose-built research workbench with native scientific rendering, multi-agent orchestration, and integrated compute management.

2. Reproducibility as a first-class feature: Every output carries an auditable history. The reviewer agent architecture directly addresses the reproducibility crisis in scientific research.

3. Real-world validation: Beta results from Manifold Bio, the Allen Institute, and UCSF demonstrate order-of-magnitude speedups on real research tasks — not just benchmark scores.

4. Anthropic as researcher, not just toolmaker: The internal drug discovery program targeting neglected diseases signals a long-term commitment that goes beyond product development.

5. Infrastructure-agnostic: Runs on your laptop, your HPC cluster, or cloud GPUs — data never leaves your systems.

6. Ecosystem play: The NVIDIA BioNeMo integration and Modal partnership create a stack that combines frontier reasoning (Claude), scientific tools (BioNeMo), and elastic compute (Modal).

---

11. Future Directions

11.1 Near-Term (2026)

• General availability — Expansion beyond beta to broader user base
• Additional domains — Beyond biology into chemistry, physics, materials science
• More specialized agents — Community-created skills and connectors
• Integration with lab automation — Connecting computational workflows to physical lab equipment

11.2 Medium-Term (2027+)

• Closed-loop discovery — Full pipeline from hypothesis to wet-lab validation to analysis
• Collaborative research — Multi-user sessions with shared agent teams
• Regulatory compliance — FDA/EMA-aligned documentation for clinical research
• Open science — Public repositories of agent-created research artifacts

11.3 The Bigger Picture

Claude Science represents a shift in how AI companies think about their role in science. Rather than providing general-purpose tools that scientists must adapt, Anthropic is building domain-specific infrastructure that understands the workflows, tools, and standards of scientific research.

This complements the agentic coding revolution documented in Claude Sonnet 5 Agentic Mid Tier Model 2026 07 03 and Openai Gpt 56 Sol Terra Luna Subagent Ultra Mode Cyber Safeguards 2026 07 06 — if those models are transforming how we build software, Claude Science is transforming how we discover knowledge.

---

References & Resources

1. Anthropic. (2026, June 30). Claude Science, an AI workbench for scientists. https://www.anthropic.com/news/claude-science-ai-workbench
2. Anthropic. (2026, June 30). Claude Science product page. https://claude.com/science
3. Anthropic. (2026, June 30). AI for Science research program application. https://docs.google.com/forms/d/e/1FAIpQLSfwDGfVg2lHJ0cc0oF_ilEnjvr_r4_paYi7VLlr5cLNXASdvA/viewform
4. Anthropic. (2026, June 30). Team plan for research labs. https://claude.com/programs/claude-team-plan-for-research-labs
5. Modal. (2026, June 30). Anthropic integration with Modal brings scalable compute to Claude Science. https://modal.com/blog/modal-integration-brings-scalable-compute-to-claude-science
6. NVIDIA. (2026, June 30). NVIDIA Announces BioNeMo Agent Toolkit. https://nvidianews.nvidia.com/news/nvidia-launches-bionemo-agent-toolkit-giving-ai-agents-the-tools-to-accelerate-scientific-discovery
7. NVIDIA BioNeMo GitHub. https://github.com/NVIDIA-BioNeMo
8. AI for Science Discourse Community. https://ai4science.discourse.group/

---

Cross-References

• Claude Sonnet 5 Agentic Mid Tier Model 2026 07 03 — Claude Sonnet 5, the agentic model powering Claude Science
• Openai Gpt 56 Sol Terra Luna Subagent Ultra Mode Cyber Safeguards 2026 07 06 — OpenAI's subagent architecture, comparable multi-agent patterns
• Claude Fable 5 Mythos 5 Redeployment Export Control Lifted Safeguards Industry Framework 2026 07 02 — Fable 5 restoration, same launch week as Claude Science
• Ai News Week 2026 06 30 2026 07 06 — Weekly roundup covering the Claude Science launch
• Gemini 3 5 Flash Agentic Frontier Multimodal Reasoning 1m Context 2026 07 03 — Google's agentic model, different approach to scientific workloads
• Qwen3 7 Max Agent Centric Era Long Horizon Execution 2026 07 01 — Qwen3.7-Max's long-horizon execution, relevant to multi-step research pipelines