Claude Science : Anthropic launches an AI workbench for researchers

Claude Science : Anthropic launches an AI workbench for researchers

🔎 Not a new model, an operating layer for science

On June 30, 2026, Anthropic unveiled Claude Science. And contrary to what the name suggests, it is not an AI model specialized in biology or chemistry.

It is a complete work environment — a workbench — that takes existing Claude models and transforms them into an operating system for scientific research. The same logic as Claude Code for developers, but applied to labs.

The timing is not coincidental. Anthropic just raised 65 billion dollars in Series H and launched Claude Opus 4.8, a sign that the company has the means to fund an infrastructure war. And for good reason: Anthropic signed with SpaceX for Colossus 1, a cluster of 220,000 GPUs and 300 MW dedicated to training Claude. Claude Science is part of this large-scale deployment strategy.

The approach is fundamentally different from what OpenAI and Google DeepMind are doing in the scientific field. And that is exactly the point.

The Essentials

• Claude Science is not a new model: it uses existing Claude models (Opus 4.8, Sonnet 4.6) via a multi-agent workflow layer.
• 60+ preconfigured connectors cover genomics, single-cell analysis, proteomics, structural biology, and chemoinformatics.
• A coordinator agent delegates to specialized sub-agents, while a reviewer agent automatically verifies citations and calculations.
• Execution is 100% local (macOS, Linux) or via SSH/HPC — sensitive data never leaves the lab.
• Reproducibility is guaranteed: each artifact includes the exact source code, environment, and full message history.
• 50 projects will be supported with $30,000 in Anthropic credits + $2,000 in Modal compute. Applications open until July 15, 2026.

Recommended Tools

What Claude Science actually does — and what it doesn't do

Claude Science is an orchestration layer. Not a model fine-tuned on scientific papers, not an AlphaFold clone.

Concretely, the tool combines a generalist coordinator agent with specialized sub-agents created by the user or preconfigured by Anthropic. The coordinator receives your request, breaks it down, assigns tasks to the appropriate agents, and compiles the results.

The artifacts produced are reproducible by design. Claude Science generates 3D protein structures, genome browser tracks, chemical drawings — but above all, it produces the exact code and message history that led to each result. According to Reuters, this approach directly targets the problem of reproducibility in science.

The crucial point: data remains local. The tool runs on your machine (macOS or Linux) or via an SSH connection to an HPC cluster. Your genomic or clinical data is never sent to Anthropic's servers. Only the calls to the Claude models go through the API — just like any other use of the Anthropic API.

This is an architectural choice that fundamentally distinguishes Claude Science from a simple web interface. And it's what allows Anthropic to target labs working with data subject to strict regulations (HIPAA, GDPR).

Multi-agent architecture: coordinator, specialists, and reviewer

The system relies on three types of agents with clearly separated roles.

The coordinator agent is the entry point. It understands your natural language request, breaks it down into subtasks, and distributes the work. It decides whether to call a genomic connector, run a Python script, or query a protein database.

The specialized sub-agents execute the tasks. Anthropic has preconfigured agents for each domain (genomics, proteomics, chemoinformatics), but users can create their own specialized agents. This is where the NVIDIA BioNeMo skills come into play — Evo 2 for genomics, Boltz-2 for proteins, OpenFold3 for structural biology.

The reviewer agent is perhaps the most strategic component. It systematically verifies the citations and calculations produced by the other agents. In a field where hallucinations can have real consequences (a poorly targeted drug, an erroneous genetic analysis), this layer of automatic verification is not a luxury — it is a necessity.

This architecture echoes what Anthropic did with Claude Code Agent View, le dashboard qui tue le terminal split-screen. Same philosophy: a main agent that orchestrates, an interface that makes the process visible, and complete traceability.

The 60+ connectors: genomics, proteomics, chemoinformatics

Connectors are the operational core of Claude Science. Without them, it would just be another scientific chatbot.

According to Anthropic's official announcement, the workbench integrates over 60 pre-configured skills and connectors, divided into five major domains.

Genomics: access to sequence databases, alignment tools, variant analysis. This is the connector that Stephen Francis (UCSF Brain Tumor Center) used to accelerate germline analysis of gliomas by a factor of 10 — a result documented by TechCrunch.

Single-cell analysis: pipelines for single-cell RNA-seq data processing, clustering, cell annotation. A field where computational complexity is a major bottleneck for many labs.

Proteomics: protein identification, quantification, post-translational analysis. Coupled with NVIDIA BioNeMo skills (Boltz-2, OpenFold3), this connector allows predicting 3D protein structures directly from the workbench.

Structural biology: visualization and manipulation of molecular structures, integrating the outputs of OpenFold3 and other prediction tools.

Chemoinformatics: according to GEN, this connector even allows designing antibiotics through a simple text prompt — an application that illustrates the tool's transformative potential in drug discovery.

Integration with Modal allows scaling compute from a single GPU to hundreds of GPUs for the heaviest analyses. This is what makes the 60+ connectors truly usable in practice, beyond just demonstrations.

Anthropic vs OpenAI vs Google DeepMind: three strategies for AI science

The launch of Claude Science reveals three radically different approaches to applying AI to scientific research. And each reflects the philosophy of its creator.

Anthropic: the open workflow. Claude Science is accessible on all paid plans (Pro, Max, Team, Enterprise). No specialized model, no domain-specific fine-tuning. Anthropic is betting on the workflow layer and multi-agent architecture to create value. It's the "broad and affordable" approach — an individual researcher in a small lab can access it just as easily as a team at Novo Nordisk.

OpenAI: the gated model. GPT-Rosalind, launched in April 2026, is a model fine-tuned specifically for biology. But it is reserved for enterprise customers with controlled access: Amgen, Moderna, Novo Nordisk, Thermo Fisher. No Pro or Max version. OpenAI is betting that the raw performance of a specialized model justifies restricted access. It's the "premium and exclusive" approach.

Google DeepMind: the owned models. AlphaFold and AlphaGenome are proprietary specialized models, coupled with Gemini for Science which integrates more than 30 databases. Google doesn't do workflows — it builds models that solve specific problems with unrivaled precision. It's the "deep and proprietary" approach.

Which of these approaches will dominate? Too early to say. But early adopters of Claude Science are delivering concrete results that make a case for Anthropic's strategy.

Concrete use cases: what early adopters have already accomplished

The documented results are impressive enough to warrant a closer look.

Allen Institute — Computational reviews of 100+ pages. Jérôme Lecoq built a multi-agent computational review pipeline with Claude Science. Result: 10 reviews of over 100 pages, with automatically verified citations. This process previously took two years. According to TechCrunch, it was reduced to a few months. The reviewer agent is central here — without it, reviews of this magnitude would be unusable.

UCSF Brain Tumor Center — Germline analysis of gliomas. Stephen Francis used Claude Science to accelerate the germline analysis of brain tumors by a factor of 10. In a field where every day counts for patients, this gain is not merely anecdotal.

Manifold Bio — Tissue-specific drug targeting. The biotech used the workbench to identify drug targets specific to certain tissues — a task that requires cross-referencing genomics, proteomics, and chemoinformatics data. Exactly the type of multi-domain workflow Claude Science is designed for.

These cases highlight something important: the value lies not in the model itself, but in the orchestration. A researcher could theoretically achieve the same thing with Python scripts and direct API calls. But the glue engineering effort required would be colossal. Claude Science eliminates this friction.

Compute and infrastructure: the role of Modal and HPC

A scientific workbench without suitable compute is a useless tool. Anthropic understood this.

Integration with Modal allows you to scale from a local GPU to hundreds of GPUs in a few clicks for analyses that require it. According to HPC Wire, this integration is a key element of Claude Science's strategy for high-performance computing.

The execution model is flexible. You can run everything locally on your Mac or Linux machine, with Python, R, and the shell directly accessible. You can also connect via SSH to your institution's HPC cluster. Or you can offload heavy computations to Modal.

This flexibility is strategic. Many labs already have institutional compute resources. Claude Science doesn't replace them — it plugs into them. And when those resources aren't enough, Modal is there as a safety net.

For the underlying models, Claude Science uses existing Claude models. If you have a Pro plan, you use the models accessible with Pro (notably Claude Sonnet 4.6, which scores 83 on the general benchmark and 81.4 on the agentic benchmark according to June 2025 rankings). If you have a Max or Enterprise plan, you access Claude Opus 4.7 (general score 90, agentic score 94.3). For a detailed comparison of models in the context of code and agentic tasks, see our Claude vs ChatGPT comparison and our guide to the best LLMs for coding.

The funding program: 50 projects, $30,000 each, deadline July 15

Anthropic isn't just launching a tool — it's funding its adoption.

50 projects will be selected, each receiving up to $30,000 in Anthropic credits and $2,000 in Modal compute. Applications are open until July 15, 2026, with results announced on July 31. Selected projects will run from September to December 2026.

According to Anthropic, this program explicitly aims to generate use cases demonstrating the value of the workbench in real research contexts. It's a total investment of $1.6 million ($1.5M in credits + $100k in Modal compute) — modest for a company that just raised $65 billion, but sufficient to create a demonstration effect.

In parallel, Anthropic is launching its own pre-clinical drug programs for neglected diseases. Eric Kauderer-Abrams, head of life sciences at Anthropic, confirmed to Reuters that the company is using Claude Science internally for these programs. It's a strong signal: Anthropic isn't just selling a tool, it's using it itself to produce research.

Reproducibility and verification: the reviewer agent in detail

The problem of reproducibility in science is not new. But AI dangerously amplifies it.

When a researcher uses an AI chatbot to analyze data and obtains a result, how do you reproduce that result? What exact steps were followed? What versions of the tools were used? Without answers to these questions, the result is scientifically invalid.

Claude Science tackles this problem head-on. Every produced artifact includes the exact source code, the execution environment description, and the complete message history between the user and the agents. This guarantees that anyone can take the same artifact and reproduce the result in a deterministic manner.

The reviewer agent adds an additional layer of verification. It systematically checks that citations in a report point to the correct sources, that numerical calculations are correct, and that conclusions are supported by the presented data. It is far from infallible — it is still an AI model checking another AI model — but it is clearly better than nothing.

In the case of the Allen Institute's computational reviews, the reviewer agent made it possible to produce 100+ page documents with a level of citation verification that would have been humanly impossible in the same timeframe.

What Claude Science means for the future of research

The potential impact goes beyond the tool itself.

Claude Science represents a paradigm shift: moving from conversational AI (ask a question, get an answer) to operational AI (describe a workflow, get a reproducible artifact). This is the difference between asking "what is single-cell analysis?" and asking "analyze this single-cell RNA-seq data, cluster the cells, annotate the cell types, and produce a reproducible report with visualizations".

This shift has profound implications for researcher training. The required skills are evolving: less manual scripting, more workflow design and critical validation of AI outputs. The researcher of the future may spend more time verifying their AI's results than producing them.

Anthropic is betting that this workflow-centric approach — rather than a model-centric one — is the right one. No specialized model for biology, no domain-specific fine-tuning. Just a good generalist model, well-orchestrated, with the right connectors and the right guardrails. If this bet pays off, it calls into question the strategy of all competitors investing in specialized models.

❌ Common mistakes

Mistake 1: Confusing Claude Science with a new model

Claude Science is not a model. It is a work environment that uses existing Claude models. If you are expecting a "science-specialized" Claude with new reasoning capabilities, you will be disappointed. The value lies in the orchestration and connectors, not in the underlying model.

Mistake 2: Thinking that scientific data is sent to the cloud

By default, Claude Science runs locally. Your genomics data, your protein sequences, your clinical data remain on your machine or your HPC cluster. Only the texts sent to the Claude models pass through the API. This is a common point of confusion, but it is crucial for regulatory compliance.

Mistake 3: Ignoring the reviewer agent

It is tempting to disable the reviewer agent to save time. This is a mistake. In a scientific context, a hallucinated citation or an erroneous calculation can invalidate months of work. The reviewer agent is a safety net, not a luxury.

Mistake 4: Comparing Claude Science to a simple chatbot with plugins

The difference is architectural. A chatbot with plugins executes tools one by one, sequentially. Claude Science orchestrates parallel agents that communicate with each other, with an integrated verification system. It is not the same category of tool.

❓ Frequently Asked Questions

Is Claude Science available for free?

No. The beta is accessible to Anthropic Pro, Max, Team, and Enterprise subscribers. The funding program offers $30,000 in credits to the 50 selected projects, but this is a call for applications, not generalized free access.

Does Claude Science replace existing scientific tools?

No, it connects to them. The 60+ connectors bridge the gap with the databases and tools researchers already use (NVIDIA BioNeMo, genomic databases, chemoinformatics tools). Claude Science is an orchestration layer, not a replacement.

Which Claude model is used in the background?

The one included with your subscription. On Pro, you will primarily use Claude Sonnet 4.6. On Max or Enterprise, Claude Opus 4.8 is available. Claude Science doesn't change the model — it changes how you use it.

How does Claude Science compare to OpenAI's GPT-Rosalind?

The approaches are opposites. GPT-Rosalind is a fine-tuned model for biology, reserved for enterprise. Claude Science is an open workflow using generalist models, accessible from the Pro plan onward. Anthropic bets on orchestration, OpenAI on model specialization.

Do the data remain confidential?

Yes, execution is local or via SSH/HPC. Scientific data are not uploaded to Anthropic's servers. Only prompts and responses pass through the API, as with any standard use of Claude.

Can you create your own specialized agents?

Yes. Anthropic provides preconfigured agents for the five covered fields, but users can create specialized agents tailored to their specific research domain.

✅ Conclusion

Claude Science is not a new AI model — it is proof that Anthropic has understood that the value in scientific AI lies not in the model, but in the workflow. By betting on an open, local, and reproducible operating layer, Anthropic is taking the opposite approach to OpenAI (gated model) and Google DeepMind (proprietary models). The results from early adopters are concrete enough to take this approach seriously. Applications for the funding program are open until July 15, 2026 on Anthropic's website.
```