Key takeaways
Claude Science is Anthropic's new AI workbench for scientists, launched in beta on June 30, 2026 — not a new model, but a workflow layer built on Anthropic's existing Claude models.
The workbench pulls together 60+ scientific databases, coding tools, and compute into one workspace, so researchers can move from literature review to a publication-ready figure without switching tools.
Every output it produces carries an auditable history — the code, environment, and conversation that created it — so results can be checked and reproduced.
Access is in beta for Claude Pro, Max, Team, and Enterprise subscribers on macOS and Linux, with discounted plans for academic and non-profit labs.
Anthropic positions the tool alongside Claude Code and Claude Cowork as one of its flagship vertical products, part of a life-sciences push the company has been building since October 2025.
Anthropic spent 2025 building coding agents. In 2026, it turned the same approach toward the lab bench. Claude Science, announced on June 30, 2026 at an Anthropic event for pharmaceutical executives, biotech founders, and researchers, is the company's attempt to do for scientific research what Claude Code already does for software engineering: take a messy, tool-fragmented workflow and put an AI agent in the middle of it. This piece walks through what the workbench actually does, what it's built on, who's already using it, and what to check before you point it at your own research.
What Is Claude Science?
It's an AI workbench designed to help scientists streamline research, analyse data, and manage complex computing workflows, according to Anthropic's own announcement. It combines databases, coding tools, compute, and research workflows into a single workspace, with the goal of letting researchers analyse literature, run analyses, create figures and manuscripts, and trace every result back to the source code and environment that produced it.
Anthropic has been explicit that this is a workflow product, not a new frontier model — it runs on the company's existing Claude model family, coordinated through a generalist agent rather than a bespoke science-specific model. Eric Kauderer-Abrams, Anthropic's head of life sciences, told MIT Technology Review that the company sees it as sitting at the same level as Claude Code and its agentic collaboration tool, Claude Cowork — internally, it's being treated as one of Anthropic's most significant product launches of the year, not a niche research add-on.
Inside the Workbench: What It Actually Does
When a researcher opens a session, they interact first with a generalist coordination agent that has access to more than 60 pre-configured skills and connectors, spanning genomics, single-cell analysis, proteomics, structural biology, and cheminformatics. Named databases reportedly include UniProt, PDB, Ensembl, ChEMBL, and ClinVar — the kind of resources scientists currently juggle by hand across browser tabs, terminals, and notebook environments like Jupyter and R.
The coordination agent can also call in specialist agents for more advanced life-sciences tasks. Anthropic has partnered with Nvidia to bring in skills from the BioNeMo Agent Toolkit, giving the workbench access to models like Evo 2, Boltz-2, and OpenFold3 without Anthropic having to rebuild that domain expertise itself — as the company put it, "the specialized science stays with the partners who built it."
The tool is also artifact-first: it can render scientific outputs natively, including 3D protein structures, genome browser tracks, and chemistry drawings, rather than treating them as generic image attachments.
Reproducibility Is the Whole Pitch
The feature Anthropic leans on hardest is traceability. Every figure or table the workbench generates carries the exact code and computing environment used to produce it, a plain-language explanation of the steps taken, and the full conversation history behind it. In practice, that means a researcher can ask it to remove gridlines or switch to a logarithmic axis in natural language, and still be able to reconstruct exactly how the final version was produced months later — reviewers and collaborators can trace, iterate, revert, or fork from any point in that history.
That's a deliberate contrast with how a lot of exploratory research code actually gets written today: quickly, in a notebook, with no real record of which version of which script produced which figure. The design makes that audit trail the default rather than something a researcher has to maintain separately.
Where It Actually Runs
The workbench is designed to run on infrastructure the lab already controls — locally on a researcher's own macOS or Linux machine, or remotely via SSH into a lab's own HPC cluster or on-demand compute account. Large jobs are planned by the coordinating agent and then submitted to that existing infrastructure, with only the context needed for each analysis step sent to Claude itself. For labs handling sensitive data — patient records, unpublished results, proprietary compounds — that's a meaningful design choice, since it keeps the bulk of the data on infrastructure the institution already controls rather than requiring it to move wholesale into a hosted environment.
Access, Pricing, and the Research Grants Program
Claude Science is currently in beta, available to any user on Claude Pro, Max, Team, or Enterprise plans on macOS and Linux — for Team and Enterprise accounts, an admin needs to switch it on first. Anthropic is also offering discounted Team-tier pricing specifically for academic and non-profit research labs, on top of the standard Claude plan tiers.
Alongside the launch, Anthropic said it would fund up to 50 research projects with as much as $30,000 in model credits each, with Modal contributing a further $2,000 in compute credits per accepted project. That grant window has already closed — applications were due July 15, 2026, with accepted projects notified by July 31 and project periods running from September 1 through December 1, 2026 — so it's worth checking Anthropic's site directly for whether a further funding round has since been announced.
Who's Actually Using It?
Early access partners named in coverage of the launch include the Allen Institute, where neuroscientists have reportedly used the workbench to shorten literature-review cycles, and Manifold Bio, which is applying it to drug-target screening. One UCSF epidemiologist was quoted describing a glioma analysis pipeline that ran in roughly a tenth of the time a conventional approach would have taken. Treat these as early, vendor-adjacent case studies rather than a guarantee of similar results on your own workload — the honest read is that this looks genuinely useful for computational biologists who aren't primarily software engineers, which is a real and underserved group, but a full independent audit of output quality across disciplines hasn't happened yet.
Why Anthropic Is Building This Now
This isn't Anthropic's first move into scientific research. In October 2025, the company released a lighter set of plug-ins under the name "Claude for Life Sciences," giving Claude access to scientific software and databases without a dedicated standalone product. Claude Science is the full-featured successor to that effort, and it arrives alongside other reporting that Anthropic hired John Jumper — a Nobel laureate and co-creator of AlphaFold — away from Google DeepMind earlier in 2026, which most coverage read as a signal of how seriously the company is taking this space.
Anthropic has also said it will use the workbench internally to pursue its own research into treatments for rare, neglected diseases — positioning the product not just as something it sells, but as a tool it's willing to bet its own research agenda on. The launch also lands as Anthropic is widely reported to be preparing for an eventual IPO, and a credible new revenue line in life sciences and pharma is a reasonable business rationale on top of the stated scientific mission.
The Competitive Picture
This product doesn't exist in a vacuum. Anthropic is one of several major AI labs courting scientific and pharmaceutical research users with agentic tooling, and coverage of the launch has widely framed it as Anthropic staking out ground in that race rather than being first to it. Given how quickly this space is moving, treat any specific competitive comparison as a snapshot rather than a settled ranking — it's worth checking current offerings directly before assuming this is the only serious option for a given workflow.
Frequently Asked Questions
Is Claude Science a new AI model?
No. Anthropic has been explicit that it is not a new frontier model — it's a workflow product built on the company's existing Claude models, coordinated by a generalist agent with access to specialized skills, connectors, and third-party life-sciences models.
Who can access it right now?
It's in beta for Claude Pro, Max, Team, and Enterprise subscribers on macOS and Linux. Team and Enterprise accounts need an admin to enable it. Academic and non-profit research labs can access discounted Team-tier pricing.
Does it send my data to Anthropic's servers?
It's designed to run on infrastructure the lab controls — locally or via a lab's own remote/HPC setup — and only sends the context needed for a given analysis step to Claude. That's a meaningfully different data-exposure profile than a fully hosted tool, but institutions handling regulated or sensitive data should still confirm the specifics against their own data-governance requirements before adopting it.
What kinds of research is it actually good at?
Public reporting and early case studies point toward computational biology, genomics, proteomics, structural biology, and cheminformatics, reflecting the 60+ pre-configured connectors it ships with. It's positioned as strongest for researchers who need to do meaningful computational work but aren't primarily software engineers themselves.
Is the research grants program still open?
Not as of this writing. Applications for the initial round (up to $30,000 in model credits per project, plus $2,000 in compute credits from Modal) closed July 15, 2026, with the funded project period running through December 1, 2026. Check Anthropic's site for whether a further round has been announced since.
If your team is evaluating Claude Science, or agentic AI tooling for research more broadly, and wants a second opinion on where it fits your actual workflow, talk to our team about a scoped AI integration review.