Claude Science

Claude Science is an artificial intelligence research workbench released by Anthropic on June 30, 2026. It is a customizable, agentic application that pulls a scientist's scattered tools, databases, and compute into one workspace, where a coordinating Claude agent can carry out multi-step research from literature review to publication-ready figures and manuscripts ^[1][3][4]. MIT Technology Review described the product as intended to support scientific research in the same way that Anthropic's coding tool Claude Code supports software engineering, carrying out meaningful work autonomously when given concise, high-level instructions ^[4]. Claude Science is not a new model. It runs the same Claude models already available to Claude users, including Claude Opus 4.8 at launch, so the bet is on workflow rather than raw model capability ^[2][3]. As of September 2026 it remains in beta; Anthropic added a Windows version in September 2026, and its own life sciences researchers have used it for protein design, model optimization and the follow-up work on the array-associated reverse transcriptase system it announced on September 23, 2026 ^{[25][27][28][29]}.

What it is

Claude Science is described by Anthropic as a single environment in which researchers can conduct all stages of their work: analyzing the literature, writing and running analysis code, querying reference databases, generating figures, and drafting papers ^[1]. Rather than a chatbot that answers questions, it behaves as a research assistant that plans and executes. Users interact with a generalist coordinating agent that has access to more than 60 curated skills and connectors, pre-configured for fields such as genomics, single-cell analysis, proteomics, structural biology, and cheminformatics ^[1][3]. The coordinating agent can hand work to specialist sub-agents and to custom specialist agents that users build for their own research ^[1][3].

The product builds directly on Claude for Life Sciences, which Anthropic launched in October 2025 to add scientific connectors, Agent Skills, and a life-sciences prompt library to Claude ^[14]. Claude Science packages that groundwork into a dedicated workbench aimed at computational biology and drug development, and Anthropic positions it as a flagship release alongside Claude Code and its agentic work assistant Claude Cowork ^[4].

Launch and positioning: "workflow, not a new model"

Anthropic unveiled Claude Science on Tuesday, June 30, 2026, at an event in San Francisco for pharmaceutical executives, biotech founders, and researchers, which Yahoo Finance identified as the company's "The Briefing: AI for Science" event ^[4][5][7]. CEO Dario Amodei spoke at the launch. According to STAT News, he said that until now humans have wrestled with the complexity of biology only with their minds, and he described AI as "a general purpose technology that helps us to make sense of that complexity." He was candid about the uncertainty, adding: "We don't know for sure if that's going to work out. But I think we're seeing signs that we're seeing the beginnings of it" ^[5].

The single most emphasized point at launch was what Claude Science is not. Anthropic stated plainly that it is "not a new AI model and not a more capable model for biology," and that it "runs the same Claude models already available to everyone today," with no special access and no gating ^[3]. TechCrunch summarized the strategy as a bet on workflow rather than a new model, reflecting Anthropic's broader shift toward vertical, workflow-level products instead of selling raw model capability ^[3]. Eric Kauderer-Abrams, Anthropic's head of life sciences, put the ambition bluntly: "It represents how important this is to our mission that this is right up there with Claude Code and Claude Cowork as the next really significant product that we're releasing" ^[4]. MIT Technology Review also reported that Anthropic would use the product to pursue its own research into drug candidates for neglected diseases ^[4].

The launch also came days after a notable hire that underlined Anthropic's science push. On June 19, 2026, John Jumper, who shared the 2024 Nobel Prize in Chemistry for AlphaFold, announced he was leaving Google DeepMind for Anthropic after "nearly 9 years" ^[15]. MIT Technology Review cited the move as a major scientific vote of confidence in Anthropic ^[4].

How it works

Skills, toolkits, and database connections

At the center of Claude Science is a library of more than 60 curated skills and connectors ^[1]; Anthropic's product page and TechCrunch describe the app as able to query more than 60 scientific databases ^[2][3]. These wire the agent into the reference resources scientists actually use, including UniProt, the Protein Data Bank (PDB), Ensembl, Reactome, ClinVar, ChEMBL, and GEO, alongside tools such as PubMed ^[1]. The toolkits cover genomics, single-cell analysis, proteomics, structural biology, and cheminformatics, and the workbench uses the skills in NVIDIA's BioNeMo Agent Toolkit to connect to models such as Evo 2, Boltz-2, and OpenFold3 ^[1]. Under the hood this extends Anthropic's Agent Skills and its connectors, which use the Model Context Protocol, the open standard Anthropic uses to connect Claude to external tools and data ^[1][14]. Labs can save any pipeline as a reusable skill or connect a preferred tool through a connector, and future sessions inherit it automatically ^[1].

Auditable artifacts and reproducibility

Anthropic stresses that "scientific research is inherently visual," and Claude Science renders scientific artifacts natively: 3D protein structures, genome-browser tracks, and chemical structures, which researchers can annotate inline and iteratively refine toward publication ^[1]. Reproducibility is a design goal. When the system produces a figure, it saves the exact code and computing environment that generated it, a plain-language description of how it was built, and the full message history, so a result can be traced back to its source months later ^[1]. A separate reviewer agent flags incorrect citations, untraceable numbers, and figures that do not match their underlying code ^[1][3]. Sessions can be forked at any point to compare two approaches ^[1].

Compute and data locality

At launch Claude Science ran locally on macOS or Linux and could reach remote resources over SSH or HPC login nodes, submitting jobs to a lab's existing infrastructure and scaling from a single GPU to hundreds as needed ^[1]. It also integrates with Modal for on-demand compute ^[1]. Because it runs on the researcher's own systems, Anthropic says large or sensitive datasets "never have to leave the systems they're already on," with only the context needed for each step sent to Claude, and the agent asks before reaching new resources ^[1][17]. Anthropic's FAQ adds that content included in prompts and model responses is processed by Anthropic under standard retention ^[2]. A Windows version followed on September 10, 2026 (app version 0.1.47); as of September 2026 the documentation lists macOS 13 or later, Windows 11 (x64), and glibc-based Linux x64 as supported systems ^[25][26].

Relationship to Claude Code and Claude's agentic stack

Anthropic repeatedly draws the analogy to Claude Code, its agentic coding tool ^[4]. Like Claude Code, Claude Science can autonomously carry out meaningful work from concise, high-level instructions, using tools and delegating to sub-agents rather than only chatting ^[1][4]. Kauderer-Abrams told MIT Technology Review that Claude Science is not meant to displace Claude Code and Claude Cowork in scientists' workflows but to build on them, for example by helping scientists run code on computer clusters ^[4]. It sits within a growing family of purpose-built Claude products that also includes Claude Code, the Claude Cowork agentic work platform, @Claude in Slack, Claude Design, and Claude Security ^[7]. The common thread is agentic behavior wrapped around the same underlying Claude models, tuned to a specific professional workflow.

Target fields and users

Claude Science is aimed first at computational biology, drug discovery, and pharmaceutical research operations, with adjacent use in genomics, proteomics, cheminformatics, and structural biology ^[4][5]. TechCrunch reported that Anthropic named Novo Nordisk and the Allen Institute as customer case studies ^[3]. Several concrete uses were shown at or around launch:

An independent hands-on account gives a sense of scale and cost. Writing in Forbes on launch day, John Drake, who studies how diseases spill over from animals into humans, handed Claude Science his research library of 6,576 papers and had it read the 490 papers on zoonotic spillover and pandemic emergence, building a bottom-up "latent ontology" of the field. He reported that of 915 relationships the literature uses repeatedly, 864 had no counterpart in the standard reference, and that the field's working language was about four times richer than its official ontologies; a figure caption put the learned ontology at 1,240 conceptual classes. Drake wrote that the project cost "twenty-six dollars and change" ^[9].

Notable uses by Anthropic

After launch, Anthropic's own science teams reported several projects run in Claude Science. The results below are Anthropic's own claims about its own work.

In the ART announcement, Anthropic introduced a new life sciences research group and laboratory and said it does its work "in Claude Science and Claude Code, the same tools available to any scientist, and sometimes with a harness of our own that coordinates many Claude sessions running in parallel" ^[29]. Its technical report describes one Claude Science session autonomously finding and retrieving RNA reads from a published infection time-course study of Staphylococcus phage SA1, which showed that ART repeat arrays are highly expressed during infection ^[30]. Anthropic said the system's function is not yet known ^[29], and TechCrunch noted that it will be up to the broader research community to judge how big or new the discovery is ^[34].

The AI for Science projects program

Alongside the launch, Anthropic said its AI for Science program, which it had started in May 2025 to give researchers free API credits ^[35], would support up to 50 Claude Science projects with up to $30,000 in credits, and that Modal would provide up to $2,000 in compute for select projects ^[1][3]. Applications were open through July 15, 2026, with award notifications by July 31, and selected projects run from September 1 to December 1, 2026 ^[1][3]. TechCrunch quoted Anthropic as seeking "postdoctoral and graduate projects that span domains," with an early focus on biomedical research ^[3].

At launch the beta was available to all paid Claude subscribers on the Pro, Max, Team, and Enterprise plans, with Team and Enterprise requiring an administrator to enable it, and Anthropic offered a discounted Team plan for academic and nonprofit research labs ^[1].

On August 27, 2026, Anthropic expanded these programs. It opened 10,000 seats for scientists to use Claude subscriptions for one year through a new Claude Team plan for scientists: standard seats are free and premium seats with 5x usage limits cost $15 per month, and principal investigators at academic or nonprofit research institutions must verify before adding their lab members. Anthropic said it intended to extend the program beyond the initial 10,000 seats. It also widened the AI for Science program beyond the biological sciences, with up to $50,000 in credits per project open to any researcher ^[24]. Anthropic's product page says the discounted plan for scientists includes access to Claude Science ^[2].

Models and safeguards

Claude Science uses whichever Claude models the user's plan includes ^[2], so it inherits their safety restrictions. In its August 27, 2026 announcement Anthropic said researchers in biology and chemistry were for now limited to its Opus-class models, because its Claude Fable models block professional biology and drug development queries ^[24]. From version 0.1.41 (August 27, 2026), the app warns users who pick a model whose safeguards block most life sciences research and offers a one-click switch, ^[25]; the safeguards documentation adds that it can be set to retry a flagged message on another model automatically ^[32].

On September 17, 2026, Anthropic introduced the Life Sciences Verification Program (LSVP) in beta, giving verified organizations access to Claude Mythos, Opus, and Sonnet models with safeguards that are more permissive for biology; Anthropic said the grants can be used through all its product surfaces, including Claude Science, and that users can switch between grants natively in the API and in Claude Science ^[31]. The Claude Science documentation says the program applies to verified organizations on Team and Enterprise plans, and version 0.1.49 (September 16, 2026) added a way to choose an approved research use case for each session ^[25][32]. When Anthropic released Claude Opus 5.5 on September 22, 2026, it deployed the model with safeguards similar to those on Claude Fable 5.1 and invited vetted organizations to apply to the LSVP to use Opus 5.5 for biology research ^[33].

Business and strategic context

Claude Science arrived as Anthropic prepared to go public. Yahoo Finance reported that the company filed confidential IPO paperwork with the Securities and Exchange Commission on June 1, 2026, that it was on track to go public later in the year, and that it was last valued at $965 billion ^[7]. MIT Technology Review reported that Anthropic said it was set to see its first profitable quarter, and noted that pharmaceutical companies have far deeper pockets than academic researchers ^[4]. Multiple outlets framed Claude Science explicitly as an effort to expand revenue streams ahead of the IPO, moving Anthropic's enterprise business beyond coding and workplace AI and into the pharmaceutical and scientific market ^[7][22].

Reception and the OpenScience open-source alternative

Coverage of the launch was broad, including MIT Technology Review, which called Claude Science Anthropic's newest flagship product, and STAT News, which described it as a product aimed at researchers and the pharmaceutical industry ^[4][5]. A comparison by Labcritics cautioned that such workbenches help most with the early, computational stages of research, that the built-in reviewers rely on the same underlying models, and that wet-lab work and peer review remain the real checks ^[17].

The most visible response came from the open-source community. Within a week of the launch, Synthetic Sciences, a San Francisco startup in Y Combinator's Winter 2026 batch, released OpenScience, an Apache-2.0 licensed, model-agnostic workbench pitched directly as an open alternative to Claude Science; MarkTechPost covered the release on July 5, 2026 ^[16][18][19]. Where Claude Science runs only on Claude models, OpenScience is designed to work with any frontier or open-weight model using the user's own API keys, including Claude, GPT, Gemini, GLM, Kimi, and DeepSeek, as well as local fine-tunes via Ollama, with switching possible per request ^[16][17][18]. At release it advertised more than 250 editable skills, and databases such as UniProt, PDB, Ensembl, ChEMBL, PubChem, arXiv, OpenAlex, and Semantic Scholar, plus around 30 more, exposed as agent tools ^[16]. The Chinese outlet QbitAI, in an article republished by 36Kr, described the skill count as more than four times that of Claude Science ^[18]. The project is independent and is not affiliated with or endorsed by Anthropic ^[16].

The comparison should be read carefully. Labcritics noted that the skill counts measure different things: OpenScience emphasizes breadth and editability, while Claude Science offers a smaller, curated, supported set with specialized integrations ^[17]. It concluded that OpenScience suits computationally fluent researchers who value vendor independence and editable tooling, while Claude Science suits bench scientists who prefer a natural-language, low-code workflow without model management; it also noted that OpenScience states its agent is not sandboxed and recommends container or VM isolation, whereas Claude Science adds an approval step before reaching new resources ^[17]. Synthetic Sciences was founded in 2025 by Aayam Bansal and Ishaan Gangwani ^[19]. The OpenScience repository was posted to Hacker News on July 4, 2026 ^[20]. As of September 2026 its README describes a desktop app for macOS, Windows and Linux and "hundreds of bundled skills" ^[23].

How it compares with related efforts

Claude Science is one entry in a fast-moving contest to build AI tooling for research, part of the broader AI for science movement. Its closest competitors take different approaches:

The sharpest contrast is with OpenAI's science work: OpenAI's GPT-Rosalind, announced April 16, 2026 and named after Rosalind Franklin, is a model series built for biology, drug discovery and translational medicine, offered through a trusted-access program to customers such as Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific ^[21]. On September 11, 2026, OpenAI said GPT-Rosalind was coming out of research preview, with published pricing taking effect on October 5, 2026 ^[21]. Google launched Gemini for Science at its I/O conference in May 2026 and pitched it as "a scientific workbench on your desktop," according to The New Stack ^[10]. Anthropic's counter-bet is that scientists do not need a new model so much as a better environment around the models that already exist ^[3][4].

Limitations and open questions

The central open question is verification. In science, a confident but wrong answer is worse than none, and large language models can hallucinate citations, numbers, and mechanisms. Claude Science addresses this with a reviewer agent that flags bad citations and mismatched numbers, but as John Drake wrote in Forbes, "It is the same model checking its own work for now, not an independent source of truth" ^[9]. TechCrunch made the same point ^[3]. Its outputs still require expert validation and, ultimately, experimental confirmation and peer review ^[17].

Reproducibility is a genuine improvement (saved code, environment, and message history make results traceable), yet reproducibility in AI for science remains a hard, unsolved problem: proprietary models, shifting weights, and stochastic outputs all complicate independent replication, a challenge the wider AI for science field continues to wrestle with. Anthropic itself frames the underlying scientific bet as unproven; Amodei conceded the company does not know for certain the approach will work ^[5]. Because Claude Science relies on existing Claude models, its ceiling is set by those models' scientific reasoning and by their safeguards: as of September 2026, Anthropic's Fable models block professional biology and drug development queries, and Claude Opus 5.5 carries similar biology safeguards, so organizations whose biology work runs into those safeguards can apply to the LSVP ^[24][31][33]. Much of its value also depends on the quality and coverage of the connectors and the skills a lab is willing to build and maintain.

ELI5 (plain-language summary)

Imagine a very capable research assistant that lives on your computer. You tell it, in plain English, what scientific question you are chasing. It reads the relevant papers, writes and runs the analysis code itself, looks things up in the big biology and chemistry databases scientists use, draws the charts and 3D molecule pictures, and drafts the write-up, then keeps a careful record of exactly how it did everything so you (or anyone else) can check and redo it later. That is Claude Science. Anthropic did not build a smarter brain for it. It uses the same Claude that anyone can already use, and gives that Claude the right tools, connections, and memory to work like a scientist. Days after it came out, a small startup released a free, open-source workbench called OpenScience that does something similar but lets you plug in any AI model you like.

See also

• AI for science
• Claude for Life Sciences
• Array-associated reverse transcriptase
• OpenAI for science
• OpenScience
• Anthropic
• Claude
• Claude Code
• Claude Opus 4.8
• Agentic AI
• Model Context Protocol
• AlphaFold
• Dario Amodei

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23Claude Science, an AI workbench for scientists - Anthropic (June 30, 2026)
2. ^1 ^2 ^3 ^4 ^5Claude Science (beta) - product page, Claude by Anthropic
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17Anthropic's Claude Science bets on workflow, not a new model, to win over scientists - TechCrunch (June 30, 2026)
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15Claude Science is Anthropic's newest flagship product - MIT Technology Review (June 30, 2026)
5. ^1 ^2 ^3 ^4 ^5Anthropic releases Claude Science, a product aimed at researchers, the pharma industry - STAT News (June 30, 2026)
6. The moment Anthropic convinced me it's serious about science - STAT News (July 1, 2026)
7. ^1 ^2 ^3 ^4Anthropic launches Claude Science in bid to expand revenue streams ahead of IPO - Yahoo Finance (June 30, 2026)
8. Anthropic Launches Claude Science Workbench For Researchers - Dataconomy (July 1, 2026)
9. ^1 ^2 ^3Anthropic's New AI Workbench Mapped My Field For $26. Now Imagine It Aimed At The Rest Of Science - Forbes / John Drake (June 30, 2026)
10. ^1 ^2Anthropic launches Claude Science, an AI workbench for scientific research - The New Stack (June 30, 2026)
11. Anthropic Launches Claude Science for Research Workflows - Enterprise DNA (2026)
12. Anthropic Launches Claude Science AI Workbench for Scientific Research - HPCwire / AIwire (June 30, 2026)
13. Anthropic launches Claude Science, an AI workspace built specifically for researchers - The Decoder (June 30, 2026)
14. ^1 ^2Claude for Life Sciences - Anthropic (October 20, 2025)
15. ^Nobel laureate John Jumper is leaving DeepMind for rival Anthropic - TechCrunch (June 20, 2026)
16. ^1 ^2 ^3 ^4 ^5Synthetic Sciences Releases OpenScience: An Open-Source, Model-Agnostic AI Workbench for Machine Learning, Biology, Physics, and Chemistry Research - MarkTechPost (July 5, 2026)
17. ^1 ^2 ^3 ^4 ^5 ^6OpenScience vs. Claude Science: Two Takes on the AI Science Workbench - Labcritics (July 7, 2026)
18. ^1 ^2 ^3100% Free Open-Source Alternative to Claude Science | Choose Between DeepSeek & GLM for Your Research Needs - QbitAI via 36Kr (July 2026)
19. ^1 ^2Synthetic Sciences: Infrastructure for Autonomous Science - Y Combinator (W26)
20. ^OpenScience: Workbench for scientific research using custom LLMs - Hacker News (July 4, 2026)
21. ^1 ^2 ^3Introducing GPT-Rosalind for life sciences research - OpenAI (April 16, 2026; updated September 11, 2026)
22. ^Anthropic launches 'Claude Science' to expand enterprise business and pharma revenue ahead of IPO - The Tech Portal (June 30, 2026)
23. ^synthetic-sciences/openscience - GitHub repository
24. ^1 ^2 ^3Expanding our support for scientists - Anthropic (August 27, 2026)
25. ^1 ^2 ^3 ^4Claude Science changelog - Claude documentation
26. ^Claude Science overview - Claude documentation
27. ^1 ^2 ^3How Claude is accelerating protein design and analytical chemistry - Anthropic (August 18, 2026)
28. ^1 ^2How Claude is uplifting biomolecular modeling - Anthropic (September 17, 2026)
29. ^1 ^2 ^3 ^4Claude discovers a novel enzyme system with CRISPR-like repeats - Anthropic (September 23, 2026)
30. ^1 ^2Autonomous AI agents discover reverse transcriptases with tandem repeat arrays - Yoon, Athukoralage, Ameisen, Kauderer-Abrams, Perry and Durrant, Anthropic technical report (September 2026)
31. ^1 ^2Introducing the Life Sciences Verification Program - Anthropic (September 17, 2026)
32. ^1 ^2Safeguards - Claude Science documentation
33. ^1 ^2Introducing Claude Opus 5.5 - Anthropic (September 22, 2026)
34. ^Anthropic says its biology lab has already found something big - TechCrunch (September 23, 2026)
35. ^Introducing Anthropic's AI for Science Program - Anthropic (May 5, 2025)