What happened
On June 30, 2026, Anthropic introduced Claude Science, an application for scientists that it describes as its largest expansion of work in life sciences. The product is intended to consolidate a research workflow that otherwise crosses many databases, file formats, data pipelines, viewers, notebooks, programming tools, and compute environments. Anthropic says the workbench can analyze literature, execute multi-step research, create detailed artifacts, and support iterative changes to figures and manuscripts.
The central product claim is traceability. Claude Science renders scientific outputs such as 3D protein structures, genome-browser tracks, and chemical structures. When it generates a figure, Anthropic says it includes the code and environment used to produce it, a plain-language explanation, and the complete message history. Users can request an edit in ordinary language, such as changing an axis to logarithmic scale, and the agent changes its own code. A reviewer agent is meant to inspect citations, calculations, untraceable numbers, and whether figures match their underlying code, flagging and correcting problems during a pipeline run.
The workbench also acts as a compute coordinator. It can run locally on macOS or Linux, on a remote machine over SSH, or from an HPC login node. For large analysis jobs, it drafts a plan and asks before reaching new resources; users can review or revoke decisions before the system writes and submits a job. It can use a laboratory's HPC cluster or a Modal account for on-demand compute, scaling from one GPU to hundreds when needed. Anthropic says the system runs on infrastructure already controlled by the laboratory, so large or sensitive datasets do not need to leave those systems; only context needed for an individual analysis step is sent to Claude.
Claude Science ships with a generalist coordinating agent, more than 60 curated skills and connectors, and support for user-created specialist agents. The initial configurations cover genomics, single-cell work, proteomics, structural biology, and cheminformatics. It connects to more than 60 scientific databases and uses skills from NVIDIA's BioNeMo Agent Toolkit to reach BioNeMo models and libraries including Evo 2, Boltz-2, and OpenFold3. Researchers can connect their own models, datasets, and pipelines, then save a pipeline as a reusable skill. Anthropic says researchers in beta have used the product for single-cell RNA sequencing, CRISPR screen design, protein structure prediction, and cheminformatics. Claude Science is available in beta to Claude Pro, Max, Team, and Enterprise users; the announcement does not specify a separate price.
What the archive adds
We have not covered this before; it opens a new thread.
Trend signal
This release supports Scientific agents need artifact evidence. Claude Science is presented not simply as a system that proposes scientific conclusions, but as one that retains code, environments, histories, citations, and review checks so a scientist can inspect and reproduce its work. The product claim remains Anthropic's claim, but the design makes evidence packages a first-class interface.
Why now
Anthropic frames the launch as the next step after life-sciences work begun the previous fall, including improvements to model capability, MCP and skill connections, and partnerships. The company says the beta has already been used by researchers for several scientific tasks and that it will refine the platform from user feedback. The timing therefore reflects an attempt to package those experiments into a broadly available workbench for eligible Claude subscribers.
Why it matters
For builders of AI systems in research organizations, the consequential choice is the operating boundary: Claude Science tries to coordinate existing tools and infrastructure rather than require all scientific work to move into a new cloud data store. Its permission step before new compute resources, support for local and HPC execution, and ability to preserve reusable pipelines as skills point to a deployment model where a lab keeps control of data and tools while delegating orchestration.
The stronger lesson is that a research agent needs more than domain connectors. A useful workflow must make its figures, numbers, code, environment, and source claims reviewable. Teams building comparable systems should treat an answer as an artifact bundle with a verification path, not as a completed chat response.
What happens next
- Anthropic will add or refine Claude Science capabilities during its beta period. (3 months, 75%). Watch for: new supported scientific workflows, connectors, or access terms.
- Research teams will test whether the reviewer agent catches failures that matter in their own pipelines. (6 months, 65%). Watch for: published user accounts of reproducibility, citation, or compute-review outcomes.
Limits and uncertainty
The announcement is a product description from Anthropic, not an independent evaluation of scientific accuracy, reproducibility, or safety. It does not provide benchmark results for the reviewer agent, error rates, limits on which workloads can run successfully, or a pricing schedule beyond eligibility for existing Claude plans. It also does not establish how much data or context is transmitted to Claude in a particular configuration, nor whether every external scientific source is current, licensed, or appropriate for a given research decision. Beta user examples are illustrative rather than controlled evidence of broad performance.
Skill to build now
Choose one bounded research workflow this week, such as generating a figure from a known dataset, and define its evidence package before automating it: input version, environment, executable code, cited sources, output figure, and a review checklist. Run the workflow once manually and once with an agentic harness. The comparison will expose whether your system can reproduce a result and whether a reviewer can detect a mismatch between a claim, a number, and the code that produced it.
Reader takeaway
Claude Science is notable because it combines scientific agents with traceable artifacts and controlled compute access. Its real test is not whether it can produce an impressive analysis, but whether a laboratory can inspect, rerun, amend, and trust the chain behind that analysis.
Open question
Can the reviewer and artifact trail provide enough dependable verification for high-stakes scientific work, or will expert review remain the limiting step as workflows become more autonomous?