Anthropic’s Claude Science: A Workflow Bet, Not a New Model

Anthropic launched Claude Science on Tuesday, an AI workbench it describes as its biggest expansion of its broader life sciences strategy since Claude for Life Sciences shipped in October 2025. The app folds over 60 research databases and tools behind one coordinating agent, with a reviewer agent flagging bad citations and untraceable numbers before output goes anywhere near a manuscript. The launch lands Anthropic squarely in a three-way race with OpenAI’s GPT-Rosalind and Google’s Gemini for Science for who owns the AI workflow inside scientific research.

Claude Science is a workflow layer on top of Anthropic’s Claude models. Anthropic runs the product on the same models everyone already has access to, including Opus 4.8.

What Claude Science Actually Does

Claude Science is a desktop application that runs locally on macOS and Linux or on a remote machine over SSH, according to Anthropic’s June 30 announcement. Researchers transition between PubMed, Jupyter, R, a cluster terminal, and more, and Anthropic’s pitch is that the new workbench folds those into one environment. A generalist coordinating agent sits at the top, with access to over 60 pre-configured skills and connectors spanning genomics, single-cell, proteomics, structural biology, and cheminformatics. That coordinator can spin up specialist sub-agents and pass work to custom assistants researchers build for their own labs. A separate reviewer agent inspects outputs at every step, flagging incorrect citations, untraceable numbers, and figures that don’t match the underlying code.

The app is in beta starting today. Claude Pro, Max, Team, and Enterprise subscribers can install it, though Team and Enterprise admins need to enable it first.

Claude Science natively renders 3D protein structures, genome browser tracks, and chemical structures, so scientists can chat with the agent about a detail and annotate the figure in place until it is publication-ready. When the agent generates a figure, it also produces the exact code and environment that produced it, along with a plain-language description and the full message history. Anthropic uses the BioNeMo Agent Toolkit from NVIDIA to connect natively to life sciences models and libraries including Evo 2, Boltz-2, and OpenFold3, and any of the lab’s own tools can be saved as a reusable skill or accessed through a connector. Claude Science runs on Opus 4.8, which also handles restricted cyber and bio queries inside Anthropic’s Fable 5 and Mythos rollout.

• UniProt
• PDB
• Ensembl
• Reactome
• ClinVar
• ChEMBL
• GEO

Anthropic lists those resources as examples of the kind of life sciences sources the workbench can query and synthesize across. Specialists sit in journals and preprint servers too, alongside domain-specific open models, and Claude Science pulls from all of those without a researcher having to navigate them one by one.

Why Anthropic Calls It a Workflow Bet, Not a New Model

Anthropic told TechCrunch that Claude Science runs on the same Claude models available to everyone today: “Not a new AI model and not a more capable model for biology. It runs the same Claude models already available to everyone today (including Claude Opus 4.8), with no special access and no gating,” the company said. The product competes on the workflow layer around the existing models. TechCrunch framed this as evidence that “Anthropic is increasingly betting its growth on vertical, workflow-level products rather than just raw model capability.” TechCrunch noted Claude Code has become the operating layer for software development, and Claude Science extends that approach to life sciences.

The framing matters because the most expensive part of AI for science is the connective tissue between the literature search, the compute cluster, the figure pipeline, and the citation check. Whoever owns that connective tissue owns the default interface scientists reach for every day. “The specialized science stays with the partners who built it,” Anthropic wrote in the announcement. Claude Science is the reasoning layer on top, and the workbench is the surface that decides which partner tools researchers reach for first. The lab’s existing pipelines can be saved as reusable skills for future sessions.

The Reproducibility Layer and Its Limits

Reproducibility is the central product promise. Every output from Claude Science carries the code, the environment, the message history, and a plain-language description of how the figure was made, so a researcher can rerun or audit it months later. The reviewer agent runs continuously during a session, comparing each generated figure against its underlying code and self-correcting when it catches a mismatch. Scientists can fork a session at any point to try a different approach without losing the original thread.

The reviewer has limits, and TechCrunch flagged the obvious one: it is still the same underlying model checking itself, not an independent source of truth. That makes Claude Science auditable in the narrow sense that every output can be traced back to a prompt and a piece of code, but it does not make the science itself independently verified. Anthropic’s case studies partly address this: Stephen Francis’s UCSF group independently validated Claude Science’s glioma analysis, and Jérôme Lecoq’s Allen Institute team built actor-critic pairs with a separate critic agent evaluating content for citation fidelity. Each pair has one agent generating content and a separate reviewer agent evaluating it, which Anthropic says the workbench supports natively.

Anthropic also designed the workbench to keep sensitive data local. Large or sensitive datasets do not leave the researcher’s laptop, Linux box, or HPC login node, the company said; only the context needed for each step of an analysis is sent to Claude. Compute runs on the lab’s own infrastructure, whether that is an on-premises HPC cluster accessed over SSH or a Modal account for on-demand GPUs, and the pipeline scales from a single GPU to hundreds as the analysis grows.

One last reproducibility pitch: the user can fork a session and try two competing approaches side by side, comparing the outputs and the code that produced each. That capability is closer to a lab notebook than to a chatbot. Future sessions inherit any saved pipeline as a reusable skill, and the lab’s preferred tools can be accessed through a connector, so the workbench builds on top of what the lab already has.

What Scientists Are Already Doing With It

Three case studies anchor the launch. Manifold Bio, which designs tissue-targeting medicines that home to a specific organ or cell type, used Claude Science to nominate targets for its latest experiments, ranking candidates against criteria Manifold had learned from its own internal proprietary data. The company said Claude Science could run the analysis end-to-end in a way a general coding assistant could not.

At the Allen Institute, neuroscientist Jérôme Lecoq built a multi-agent “computational review template” with about 20 custom skills for writing long-form reviews. Sub-agents read through thousands of papers, pulled each paper’s central claim and key quantitative finding, and stored them in an evidence database. The pipeline then wrote the review section by section, with one agent creating content and a separate reviewer agent evaluating it for accuracy and citation fidelity. Before Claude Science, Lecoq’s team could spend as many as two years on a single review; the team now has about 10 reviews, many of them more than 100 pages, with citations checked over by reviewer agents.

At the UCSF Brain Tumor Center, associate professor and epidemiologist Stephen Francis used Claude Science to support molecular epidemiology studies on glioma. The work, which examines how thousands of small-effect germline variants combine to shape individual susceptibility to the brain tumor, was already underway when the workbench arrived. Francis said the app enabled comprehensive germline workups across multiple approaches in about one-tenth the time the previous process took, and his group independently validated the results.

• Over 60 skills and connectors pre-configured for life sciences
• About 20 custom skills in Lecoq’s review template
• About 10 reviews written with Claude Science, many over 100 pages
• Comprehensive germline workups in about one-tenth the previous time
• Up to two years to write a single review before Claude Science

OpenAI Went Narrow, Google Owns the Models

Anthropic is not alone in the space. In April 2026, OpenAI shipped GPT-Rosalind, a model fine-tuned for biological reasoning and launched as a research preview limited to qualified enterprise customers in the United States, with qualification and safety review gating access. Early partners included Amgen, the Allen Institute, Moderna, Thermo Fisher, and Novo Nordisk. At Google I/O in May 2026, the company rolled out Gemini for Science, also pitched as a desktop scientific workbench, bundling 30-plus life-science databases with Google’s AI co-scientist hypothesis engine.

The structural difference is what each company can claim as its own. Google DeepMind owns foundational science models like AlphaFold and AlphaGenome, which the other vendors can call into as tools but cannot match on the model’s own merits. OpenAI’s distribution play is narrow and gated, betting that qualified enterprise labs will pay for a model that knows biology well. Anthropic’s distribution is the wide one: any Pro or Max subscriber can install Claude Science today, no qualification review, and the product runs on top of the same Claude models everyone else already has.

Three Distribution Strategies in One Scientific Market

The three approaches are not converging. TechCrunch framed the contrast directly: “three very different distribution strategies are now competing for the same scientific research market: Anthropic is going wide with broad subscription access, OpenAI is going narrow and enterprise-gated, and Google is leaning on owned, proprietary models nobody else has.” Each strategy is a hypothesis about how AI vendors will compete in specialized verticals more broadly.

Anthropic’s hypothesis is that the workflow layer is the moat. If Claude Science becomes the default place scientists do their work, the underlying model becomes more interchangeable and Anthropic’s position becomes harder to dislodge. OpenAI’s hypothesis is that a qualified, safety-gated biology model sells itself to enterprise labs that need a credible specialized answer. Google’s hypothesis is that owning AlphaFold-class models is the moat, and a workbench is just the distribution surface for them.

The OpenAI path also showed the cost of the narrow strategy. The New Stack reported that OpenAI disbanded its OpenAI for Science team and sunset Prism, its scientific writing tool, after Kevin Weil’s departure in April 2026. The two moves, GPT-Rosalind’s gated launch and the closure of the broader science program, suggest OpenAI chose to focus on its core products rather than carry a full scientific software line. Anthropic went the opposite direction with a wide workflow product at the same time OpenAI narrowed.

The same underlying-model critique applies to the reviewer agent inside Claude Science, but Anthropic at least separates the reviewer from the generating agent and runs it continuously during a session. The reviewer is still the same model in the loop, so the audit trail is internal rather than external.

What This Signals for Law, Finance, and Engineering

TechCrunch’s read on the launch is that the scientific market is an early test case for how AI vendors will compete in other specialized verticals. The outlet wrote that “how that plays out could be an early signal for how AI vendors compete in other specialized verticals like law, finance, and engineering, down the line.” Anthropic’s life sciences push began in October 2025 with Anthropic’s Claude for Life Sciences chatbot, an augmentation of the Claude chatbot for life sciences tasks. Claude Science, with its desktop app and dedicated reviewer agent, is a standalone desktop application, while Claude for Life Sciences was a chatbot augmentation. TechCrunch noted Claude Code has become the operating layer for software development, and Claude Science extends that approach to life sciences.

Anthropic is also using Claude Science to recruit the next wave of users. The company said it will support up to 50 AI for Science projects with up to $30,000 in credits per project, plus up to $2,000 in Modal compute. Applications close July 15, 2026, projects run September 1 to December 1, 2026, and Anthropic framed the program as a way to learn how the product holds up across domains, starting with biology and biomedical research.

The launch is also a clean test of whether owning the workflow layer matters more than owning the best model in a given vertical. If Anthropic’s wide, workflow-first approach pulls in the same enterprise labs that OpenAI is gating access for, the narrow strategy starts to look expensive; if Google’s owned models keep winning the highest-profile scientific questions, the workflow layer becomes a thinner moat.

• Up to 50 AI for Science projects funded by Anthropic
• Up to $30,000 in credits per project from Anthropic
• Up to $2,000 in Modal compute per project
• Applications close July 15, 2026
• Projects run September 1 to December 1, 2026