Anthropic's newest flagship product is not a model. On June 30, 2026, the company launched Claude Science, an AI workspace built for researchers that runs on the Claude models it already ships rather than on some new brain trained for biology. That choice is the whole strategy. Anthropic is betting that the thing standing between scientists and useful AI is not raw model intelligence but the plumbing around it: the databases, the reproducibility, the verification, and the ability to keep sensitive data on a lab's own machines. Claude Science is an attempt to sell the plumbing.
- Claude Science is a workflow product, not a new model, and it runs on existing Claude models including Opus 4.8.
- It uses a multi-agent design where a lead assistant delegates to sub-agents, with a separate agent that fact-checks citations and calculations.
- It connects to more than 60 scientific databases and can run locally, keeping sensitive research data inside a lab's own infrastructure.
- Early users include the Allen Institute, a UCSF brain-tumor group, and Novo Nordisk, and Anthropic is funding up to 50 research projects with $30,000 in credits each.
What Claude Science actually is
Anthropic positions Claude Science alongside Claude Code and Claude Cowork as a major product line rather than a feature, and the framing matters. Where Claude Code is an operating layer for software development, Claude Science is meant to be the equivalent layer for scientific research. The company is explicit that this is not a new AI model and not a more capable model for biology. It leans on existing Claude models, Opus 4.8 among them, and wraps them in tooling designed for the specific shape of research work.
The reporting from TechCrunch makes the bet plain in its headline: Claude Science wagers on workflow, not a new model, to win over scientists. That is a notable stance in a field where competitors have reached for specialized, fine-tuned models. Anthropic's argument is that a capable general model plus the right scaffolding beats a narrow model with none, at least for the messy, multi-step reality of how research actually gets done.
What the workspace does, in practice, spans the arc of a research project. A scientist can point it at the literature to analyze prior work, run multi-step analyses on data, generate visualizations and charts, and draft manuscript sections. It can go further and propose new drug candidates, which Anthropic demonstrated with work on phenylketonuria, a metabolic disorder. The pitch is not a chatbot that answers questions but an agent that executes meaningful chunks of a project from a high-level instruction.
The multi-agent workflow at the core of Claude Science
The architecture is where the product gets interesting. Claude Science runs a multi-agent setup in which one main assistant behaves like a project manager, breaking a task down and spinning up sub-assistants to handle the pieces. That delegation is how the system tackles a job too large for a single context window or a single reasoning pass, mirroring the way a principal investigator hands work to members of a lab.
Bolted onto that is a separate verification layer. A distinct agent checks citations and calculations on its own before results are presented, which is a direct response to the well-known failure mode of language models inventing references or fumbling arithmetic. In a research setting, a fabricated citation is not a minor embarrassment, it is a poison that can propagate into published work. Building the fact-checker as its own agent, rather than trusting the main model to police itself, is a design choice aimed squarely at making the output trustworthy enough to use.
The verification agent is the feature that most distinguishes Claude Science from a general chatbot pointed at a science problem. A model that checks its own citations against a database before showing them addresses the single biggest reason researchers have been wary of trusting AI output in their work.
Reproducibility gets similar attention. When Claude Science generates a figure, whether a 3D protein structure or a chemistry diagram, it embeds the code that produced it, a plain-language description of what it shows, and the full message history that led there. A scientist can therefore trace any result back to its inputs and rerun it, which is the baseline requirement for research to count as legitimate. Better still, figures can be edited in plain language: a researcher describes the change they want, and the agent updates the underlying code to match, keeping the visual and its source in sync.
Keeping the data in the lab
The infrastructure story is as important as the intelligence story, because research data is often the most sensitive asset an institution holds. According to The Decoder, the application runs locally on macOS or Linux and connects out to remote machines over SSH or to HPC clusters, which lets a lab keep its raw data inside its own systems. Only the context the model actually needs is transmitted, and compute can scale from a single GPU up to hundreds depending on the job.
That local-first design, with an on-premises option to run against a lab's own infrastructure rather than Anthropic's servers, is a direct answer to the objection that has kept regulated and proprietary research away from cloud AI. A pharmaceutical company sitting on unpublished compound data, or a hospital group with patient genomics, cannot ship that material to a third-party API. By keeping the data local and sending only slices of context to the model, Claude Science tries to make the compliance problem tractable rather than asking institutions to trust a black box with their crown jewels.
The toolkit surface backs up the research focus. Claude Science ships with prebuilt skills covering genomics, proteomics, and cheminformatics, and it interfaces with tools for genetics and chemistry, along with protein biology. It also folds in Nvidia's BioNeMo agent toolkit, which brings specialized models such as Evo 2, Boltz-2, and OpenFold3 into reach. So while the reasoning core is a general Claude model, the workspace hands it purpose-built instruments and more than 60 scientific databases to work against, which is the difference between a smart generalist and a smart generalist equipped for a lab bench.
Who is already using it
Anthropic launched with named early adopters rather than hypotheticals, which lends the release some weight. The Allen Institute neuroscientist Jerome Lecoq built a multi-agent computational review pipeline on the platform. At the UCSF Brain Tumor Center, the group led by Stephen Francis used it to accelerate germline analysis, the study of inherited genetic variation. Novo Nordisk, the pharmaceutical giant behind a generation of metabolic drugs, appears as a named customer case study alongside the Allen Institute.
Anthropic is also eating its own cooking. The company says it is using Claude Science internally to pursue drug candidates for neglected diseases, the underfunded conditions that rarely attract commercial research budgets. That internal use is both a product demo and a mission statement, and it dovetails with how the company talks about its purpose in the space.
How capable is the underlying model at real science? One data point comes from MIT Technology Review, which reported that Harvard physicist Matthew Schwartz estimated Claude, running on an Opus model, performed at the level of a second-year graduate student when executing scientific projects. That framing is worth holding onto because it captures the current reality: useful and fast, genuinely helpful at execution, and not yet an independent principal investigator. A capable second-year graduate student can do enormous amounts of real work under direction, and can also make confident mistakes that need a supervisor's eye.
How Claude Science reaches researchers
Access is arriving through a beta available to Pro and Max subscribers, along with Team and Enterprise accounts, which puts the workspace in front of both individual researchers and institutions. MIT Technology Review noted that the product is accessible to paid Claude subscribers, so the distribution rides on Anthropic's existing subscription base rather than a separate procurement process, at least for the beta.
Anthropic is also priming the pump with money. The company set up a research grant program funding up to 50 projects with $30,000 in Claude credits each, with applications that closed on July 15, 2026 and funded projects running from September through December 2026. Seeding real projects with credits is a familiar platform tactic, and it is a sensible one here: the fastest way to prove a research workspace works is to get it into the hands of working scientists on real problems and let the case studies write themselves. The grant structure also quietly filters for serious users, since applying for and running a funded project is a higher bar than clicking through a free trial.
The competitive field Claude Science enters
Anthropic is not the only lab courting scientists, and the contrast in approaches is the clearest way to read its strategy. OpenAI took the opposite path with GPT-Rosalind, released in April 2026, a specialized fine-tuned model gated behind enterprise access and limited to qualified US customers. That is the narrow-model bet: train a system specifically for the domain and restrict who can touch it. Google DeepMind, meanwhile, owns some of the foundational models of computational biology in AlphaFold and AlphaGenome, and bundles more than 30 life science databases into its Gemini for Science platform.
Set against those two, Anthropic's play is the workflow-first, model-agnostic middle. It concedes that DeepMind may hold stronger specialized foundation models and that OpenAI may have a domain-tuned system, and it argues that neither matters as much as the operating layer a scientist actually works in day to day. The wager is that researchers will choose the environment that connects to their databases, keeps their data local, verifies its own outputs, and produces reproducible figures, even if the model underneath is a generalist rather than a biology specialist. Whether that is right is the open question the next year will answer.
There is history behind the launch, too. Anthropic started down this road with Claude for Life Sciences in October 2025, so Claude Science is the maturation of a nearly year-long push rather than a standing start. The progression from a life-sciences offering to a full research workspace tracks the broader industry move away from selling model access and toward selling the scaffolding that makes a model useful for a specific kind of work.
Why reproducibility is the hard part
It is easy to underrate how much of Claude Science is aimed at a single unglamorous problem: making an AI's work checkable. Science runs on the principle that a result means nothing unless another person can retrace the steps and get the same answer. A model that produces a beautiful figure and a confident conclusion, with no way to see how it got there, is worse than useless in that setting, because it invites a researcher to trust an output they cannot audit. Anthropic's response is to attach the machinery of an audit to every artifact the system produces.
That is why a generated figure carries its own code, a plain-language account of what it depicts, and the message history that produced it. A reviewer can open the figure, read the code that drew it, and rerun that code to confirm the numbers, the same way they would interrogate a colleague's analysis script. The natural-language editing loop reinforces the point: because a plain-language edit updates the underlying code rather than just repainting pixels, the figure and its source never drift apart. The result stays reproducible even after a scientist has reshaped it several times.
The verification agent extends the same philosophy to text. Instead of assuming the main model got its citations and arithmetic right, a second agent independently checks them against the connected databases before anything reaches the researcher. It is a belt-and-suspenders design, and it reflects a hard-won lesson about language models in high-stakes domains: the cheapest place to catch a fabricated reference is before it is ever shown, not after it has been pasted into a paper. None of this makes the system infallible, but it moves the failure rate toward the range where a careful human reviewer can manage what slips through.
The stakes for drug discovery
The domain Anthropic keeps returning to is drug development, and the phenylketonuria demonstration shows why. Phenylketonuria is an inherited metabolic disorder, exactly the kind of well-characterized genetic condition where an AI that can reason across genomics, protein structure, and chemistry might surface a candidate a human team would take longer to find. By showing Claude Science identifying new drug candidates for such a condition, Anthropic is making a concrete claim: the workspace is not just a writing aid but a tool that can participate in the discovery itself.
The named customers point the same direction. Novo Nordisk sits at the center of modern metabolic-disease drug development, and the UCSF brain-tumor group's use of the system to speed germline analysis touches directly on cancer research. These are not marketing-friendly toy problems but the working edge of biomedical science, which is the audience Anthropic most wants to convince. The company's own internal use of Claude Science to chase treatments for neglected diseases, the conditions that commercial pipelines routinely skip, rounds out the message that the tool is meant for real discovery under real constraints rather than demos.
What the workflow bet says about AI's next phase
The most telling thing about Claude Science is what it declines to be. In a market that has trained everyone to expect each announcement to be a bigger, smarter model, Anthropic shipped a product whose headline claim is explicitly that it contains no new model at all. That is a signal about where the frontier labs think the value is moving. When the base models are already strong enough to work at a graduate-student level across many domains, the bottleneck stops being intelligence and becomes integration: getting the model connected to the right data, keeping that data safe, checking the model's work, and making its output reproducible.
For scientists, the practical read is cautiously optimistic. A tool that automates literature review, runs analyses, drafts figures with traceable code, and checks its own citations could compress the tedious middle of a research project, freeing time for the parts that need human judgment. The caveats are equally practical. A second-year-graduate-student model will still make errors that a domain expert has to catch, the verification agent reduces but does not eliminate the risk of bad references, and the local-infrastructure requirements mean adoption depends on a lab's ability to run and connect the software. Claude Science is not a machine that does science on its own, and Anthropic is careful not to claim it is. It is a workspace that makes a capable assistant genuinely useful inside the constraints real research operates under, and that more modest promise may prove more durable than another round of benchmark records.