What Happened
On June 30, Anthropic launched Claude Science — an AI workbench for scientific research. It is Claude's productization for biology/chemistry/research verticals, positioned as "Claude Code for Scientists."
Core mechanisms:
- Multi-agent orchestration: Users interact with a general coordinator agent holding 60+ preconfigured skills and connectors (genomics, single-cell, proteomics, structural biology, cheminformatics). It can spin up sub-agents on demand and plug into user-built specialist agents.
- Reviewer agent: Checks citations, computations, and figures against the underlying code in real time, self-correcting discovered errors.
- Integrated scientific databases: UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, GEO and 60+ others, connected via the NVIDIA BioNeMo Agent Toolkit directly into life sciences models/libraries (Evo 2, Boltz-2, OpenFold3).
- Compute orchestration: Automatically manages HPC over SSH and Modal on-demand GPUs; scales from a single card to hundreds; sensitive data stays local, with HPC jobs transmitting only "the minimal context needed for the current step."
- Reproducible artifacts: Every figure/manuscript carries "the code + environment + full message history that created it," reproducible months later.
- Local deployment: macOS/Linux local + remote HPC + SSH/HPC login node, with forked sessions for comparing analysis paths.
- Manifold Bio (tissue-targeting drug company): used Claude Science to evaluate each candidate target's surface expression, trafficking, and safety, ranked automatically on internal proprietary data — compressing evaluations that previously took two weeks of manual work into a few conversations.
- Jérôme Lecoq, Allen Institute (neuroscientist): built a multi-agent "computational review template" with ~20 custom skills — sub-agents read thousands of papers, extract central claims and quantitative findings into an evidence state database, then write reviews section by section, each with an actor-critic pair (writer agent + reviewer agent). Reviews that used to take 2 years now number about 10 manuscripts of 100+ pages.
- Stephen Francis, UCSF Brain Tumor Center (epidemiologist): performed molecular epidemiology analysis of gliomas with a comprehensive germline workup across multiple methods — tasks that took 10x longer now complete in one pass, robust after independent validation.
- OpenAI Deep Research is another AI for Science player, but leans toward "research synthesis" rather than "experiment execution"; Claude Science extends to "execution + HPC scheduling," covering more ground.
- Google Scholar + Gemini + Vertex AI is a scattered toolchain without a unified workbench.
- Anthropic's moat is the product experience Claude Code accumulated since June — multi-agent orchestration, tool use, MCP connector ecosystem — directly ported, not built from scratch.
- Researchers: Apply for the beta now and hand the most painful parts of your workflow (review writing, literature search, parameter sweeps) to Claude Science.
- Bio/pharma CTOs: Manifold Bio proves the ROI — an evaluation going from 2 weeks to a few conversations means the same team can run 5–10x more projects.
- AI for Science founders: Anthropic has stamped out an open "agent + skill + connector" template — a product prototype for any vertical AI for Science company.
- AI coding tool companies: The vertical workbench template is out — should your Cursor / Cline / Devin ship "finance," "legal," or "education" editions?
- HPC centers: If Claude Science's orchestration model spreads, HPC usage shifts from "submit and queue jobs" to "AI agent scheduling + auto-scaling."
- Data security boundary: "Sensitive data stays local, HPC gets only minimal context" — but what defines "minimal context"? A query needing protein sequence + mutation site + expression level is itself sensitive data, not "sensitive" in the model-weights sense.
- Amplified hallucination risk: In coding, a wrong error means a compile failure at worst; in research, a wrong AI-predicted protein structure can cause million-dollar experiment failures. What's the reviewer agent's coverage? Which scenarios require human double-checks?
- Database subscription costs: UniProt and PDB are free, but some ChEMBL/ClinVar advanced features require subscriptions — does Claude Science billing cover database fees?
- Cross-institution collaboration: Academic data is usually shared, but sessions are per-user. How do multi-lab joint projects collaborate in Claude Science?
- Long-term memory and versioning: Projects run 2–5 years. Does Claude Science support multi-year sessions? How does evidence state carry over after session forks?
- NVIDIA BioNeMo lock-in: Heavy reliance on the NVIDIA BioNeMo Agent Toolkit means Anthropic must build or substitute if NVIDIA changes licensing or pricing.
- Anthropic announcement: https://www.anthropic.com/news/claude-science-ai-workbench
- Anthropic product page: http://claude.com/science
- TechTimes beta details: https://www.techtimes.com/articles/319439/20260701/anthropic-launches-claude-science-ai-research-workbench-open-all-paid-subscribers.htm
- Tencent News coverage: https://news.qq.com/rain/a/20260701A05E8U00
- Sina Finance: https://finance.sina.com.cn/tech/roll/2026-07-01/doc-inifhmie3179147.shtml
- Anthropic academic Team Plan: https://claude.com/programs/claude-team-plan-for-research-labs
- Modal × Claude Science integration: https://modal.com/blog/modal-integration-brings-scalable-compute-to-claude-science
Pricing and availability: beta open to all Pro/Max/Team/Enterprise paid subscribers, no enterprise vetting required; the Team plan offers discounted seats for academic labs/non-profits. Anthropic is also supporting up to 50 AI for Science projects, each with up to $30K in credits, plus up to $2K in Modal compute. Applications close July 15, 2026; execution runs September 1 – December 1.
Three validated use cases:
Deep Analysis
The strategic significance of Claude Science is not "researchers can finally use AI" — that happened long ago. It's that Anthropic has, for the first time, carried the "AI coding tool" product paradigm into a non-software domain:
Layer 1: Product template reuse from "prompt engineering" to "agent + skill + connector." Claude Code's core abstraction is agent + tool + skill. Claude Science reuses this wholesale, swapping tools from "read files, write code, run tests" to "query UniProt, call Boltz-2 for protein structure, submit HPC jobs." Anthropic extends "AI writes my code" to "AI does my research" with the same product language — a product architecture lever, not a new system.
Layer 2: Research is a natural extension of the AI coding paradigm. Research workflows resemble software engineering closely: literature review ≈ reading code, experiments ≈ running tests, paper writing ≈ writing docs, job submission ≈ git push. The full capability set Claude Code honed in software — agents that read context, write code, run commands, review results — carries directly into research.
Layer 3: Compute scheduling is the hard part of "AI for Science." AI coding tools face relatively simple scheduling (local + cloud GPU + API calls). Research is far more complex: HPC over SSH + local sensitive data + scaling to hundreds of GPUs + multi-step pipeline dependencies + GPU-hour billing. Claude Science invests heavily in compute orchestration: drafting plans, asking before provisioning new resources, letting users review/revoke decisions anytime, and auto-scaling from one to hundreds of GPUs. Anthropic is signaling it can do infrastructure, not just API calls.
Layer 4: Reproducibility — research's chronic pain — is directly addressed. For 20 years, the biggest research pain point has been irreproducible results: opaque analysis, unrecorded parameters, mismatched environments. Claude Science mandates that every figure carries the code, environment, and message history that created it, aligning with NIH's 2023 Data Management and Sharing Policy — AI workflows make reproducibility a default behavior rather than relying on scientists' discipline.
Impact on OpenAI and Google:
Reverse impact on the AI coding ecosystem: Claude Science provides a "vertical industry workbench" template for every AI coding tool — Cursor for iOS (mobile IDE), Cloudflare Sandbox (cloud agent runtime), Anthropic Claude Science (research IDE). General-purpose AI coding tools are being carved up into verticals; every industry may get its own Claude Code variant.