What Is Claude Science? Anthropic's AI Workbench for Researchers — How It Works and Who Can Use It (October 2026)

Claude Science is a desktop app from Anthropic that lets Claude carry out computational research end to end: search the literature, pull data from scientific databases, write and run Python or R code on your machine, push heavy jobs to a cluster, then produce figures and draft the methods section of your paper. Every result is saved with the code and environment that produced it, so someone else can reproduce it. Anthropic launched it on 30 June 2026 as a beta and positioned it plainly: what Claude Code is for developers, Claude Science is for researchers.

The most common misunderstanding is that it is a new model. It is not. Claude Science runs the same Claude models everyone already has (at launch that meant Claude Opus 4.8). What is new is the "workbench" that brings tools, databases and compute into one place. This article answers three questions: what it is, who can use it, and how to use it step by step, and closes with what a Thai research organization needs to prepare before switching it on.

1. How Claude Science differs from the Claude you already use

If you already use Claude on the web or in the Desktop app, the first question is why there is another app. The answer comes down to "who runs the code" and "how results are kept". This table compares the three routes Anthropic offers today.

Source: anthropic.com/news/claude-science-ai-workbench and claude.com/docs/claude-science/overview (16 Oct 2026)

Claude Science did not appear out of nowhere. It builds on Claude for Life Sciences, which Anthropic released in October 2025 as a set of connectors to Benchling, PubMed and biology databases inside the chat window. Here is the timeline worth knowing.

2. What is inside: a coordinating agent, 60+ skills, and a reviewer

When you open the app you are not talking to a "model" directly. You talk to a coordinating agent that takes your request, drafts a plan and hands tasks to specialist sub-agents, one that knows genomics, another that knows protein structures, and so on. The part people overlook is the reviewer, which works in the background and checks whether the conclusions Claude writes match the code that actually ran, and whether the citations it uses exist.

The official docs also stress that Claude Science is not limited to the life sciences. Anthropic lists computer science and machine learning, engineering, materials science, mathematics, physics and the social sciences as supported, because the core loop is the same in every field: form a hypothesis, review the literature, design an experiment, analyze a large dataset, and turn it into reproducible figures and findings. Only some connectors and skills are biology-specific.

3. Who can use it and what you need

The good news is that there is nothing extra to buy. Claude Science is included in every paid individual and organization plan. There are, however, machine and admin conditions to check first.

4. How to use it: six steps from install to manuscript

The steps below follow Anthropic's Get started documentation, in the order a real user meets them.

1. Check access first. On a Team or Enterprise account, an Owner opens the admin settings and enables Claude Science for the organization. On Pro or Max, skip this step.
2. Download and install from claude.com/science, choosing the installer for your machine (Mac Apple silicon, Mac Intel, Windows, Linux). The app is separate from Claude Desktop and can sit alongside it. Sign in with your existing Claude account.
3. First-time setup. In Settings, switch on the skills that match your work, add compute environments (your own machine, SSH to a workstation or a cluster login node, a Modal account), and store credentials for licensed journals or cloud storage under Settings > Credentials, where they are encrypted on your computer.
4. Open a Project and start a Session. Type your task in plain language. Claude drafts a plan and asks for approval every time before reading a new folder, contacting a new host or submitting a remote job. You can review, change or revoke before anything runs.
5. Iterate on results. Ask for figure changes in plain language and Claude edits the code and re-runs it. For heavy work, Claude drafts the job script (for example a Slurm sbatch file), shows it to you before submitting, then monitors the job.
6. Draft the manuscript. Ask for a methods section written from the pipeline that actually ran and a results section from the figures. The reviewer checks claims and citations in the background. Export the artifacts and conversation history to keep alongside the data.

A prompt you can use to start a session (change the file names and questions to match your work):

Read counts.h5ad in ~/lab/scrna-run42
Run standard single-cell QC (drop cells with fewer than 200 genes or more than 15% mito)
Cluster with Leiden and find marker genes for each cluster
Produce a UMAP coloured by cluster and a top-10 marker table per cluster
Before running anything that takes more than 10 minutes, tell me and offer to send it to the cluster

Note the last line. Stating your time and compute constraints up front lets the coordinating agent plan correctly from the start instead of running on a laptop until it stalls.

5. Where the data lives, security, and what it cannot do

The first question from any lab head or IT team is where the raw data goes. Anthropic's documentation is direct: your files stay on your computer, code runs in a sandbox, and Claude receives only the context needed for the task, not the whole dataset. Each new folder, host or remote job requires approval, and on Team and Enterprise the organization's admin decides whether members may use SSH hosts and Modal at all. For how each plan treats your data for training, see our article on Claude data governance and security.

One piece of news deserves caution. In early October 2026 Anthropic announced that Claude agents had made their "first scientific discovery", a new enzyme system and a pattern in viral DNA resembling the CRISPR (Clustered Regularly Interspaced Short Palindromic Repeats) mechanism. CNN reported that many scientists are not convinced it counts as a discovery. For everyday users this does not change the product's value, which lies in shorter analysis time and reproducibility, not in having an AI discover things on your behalf.

6. Real-world uses Anthropic has published

• UCSF Brain Tumor Center (University of California, San Francisco) used it for germline epidemiology analysis. Anthropic reports it took roughly one tenth of the previous time.
• Allen Institute built multi-agent computational review templates for its researchers to reuse.
• Manifold Bio used it for target nomination for tissue-targeting medicines.
• Novo Nordisk was among the first customers Anthropic cited at launch, according to TechCrunch.

The "one tenth" figure is self-reported by Anthropic and its customer, with no independent measurement yet. But the pattern of work that benefits is consistent: data analysis with a clear pipeline, repeated many times, needing figures with provenance for publication.

7. Compared with OpenAI's GPT-Rosalind

The three big labs chose different paths. OpenAI released GPT-Rosalind in April 2026 as a biology-tuned model gated to vetted enterprise customers. Google DeepMind relies on specialist models such as AlphaFold. Anthropic built no new model; it built a workbench on top of its general models and opened it to every paid plan. The upside of Anthropic's route is easy access and cross-field use. The limit is that biological capability is tied to the general models and to the model restrictions discussed in section 3. If your lab truly needs a domain-tuned model, compare both routes before deciding.

8. How a Thai research organization gets started

Universities, research institutes under the Ministry of Higher Education, Science, Research and Innovation (MHESI), and private-sector labs can all reach Claude Science because Thailand is a supported country. What differs is the purchase route and the internal preparation.

Whichever route you take, settle three things internally before switching it on. First, which datasets may go in: patient data or personal data must pass your organization's PDPA (Personal Data Protection Act) process first. Second, who approves SSH access to the institute's cluster and who pays for Modal. Third, a standard for storing artifacts and conversation history alongside the research data, so you can tell an auditor or a reviewer where every figure came from.

Grand Linux's role here is straightforward. We procure Claude Team and Enterprise for organizations in Thailand, issue quotations in Thai Baht and Thai tax invoices, and accept purchase orders (POs) under procurement rules (see our Claude procurement service). For organizations that also want Claude to read internal systems such as an ERP or intranet over MCP with role-based permissions and an audit trail, we build that part too; see Claude Solutions. If your organization is interested, submit a quote request and our sales team will help you work out which route in the table above fits.

Conclusion

• Claude Science is a desktop app for researchers, launched 30 June 2026 in beta. It is not a new model but a workbench that combines a coordinating agent, 60+ skills and connectors, local Python/R kernels, HPC/GPU job submission and reproducible artifacts.
• It is included in Pro, Max, Team and Enterprise at no extra cost. On Team/Enterprise an admin must enable it, and usage counts against the same limits as Claude Code and Cowork.
• Data stays on your machine, code runs in a sandbox, and every new resource needs approval, but an internet connection is required and the reviewer does not re-run analyses, so humans still verify.
• It is not a clinical tool, and biology and chemistry researchers are limited to Opus-class models.
• Thai organizations can use every route, from organization Team/Enterprise to the scientists' seat program and AI for Science credits. What to prepare: PDPA data rules, cluster access rights and an artifact retention standard.

AI does not make research more correct by itself. What Claude Science gives you is a figure that traces back to its code and data, and that is the first thing a reviewer asks for.

- Saeree ERP team, Grand Linux Solution

References

• Anthropic — Claude Science, an AI workbench for scientists (30 June 2026)
• Claude Science documentation — Overview, requirements, plans and usage (claude.com/docs)
• claude.com/science — downloads and product overview
• Anthropic — Expanding our support for scientists (27 August 2026)
• Anthropic — Building on our commitment to American scientific discovery (8 October 2026)
• Anthropic — Claude for Life Sciences (October 2025)
• TechCrunch — Anthropic's Claude Science bets on workflow, not a new model, to win over scientists (30 June 2026)
• MIT Technology Review — Claude Science is Anthropic's newest flagship product (30 June 2026)
• CNN — What to know about Anthropic's 'Claude-led' biological discovery, and why scientists aren't convinced (8 October 2026)