# 🔬 Claude Science — Investigation Summary ## What it is **Claude Science** is an AI workbench for scientists, launched by Anthropic on **June 30, 2026**. Critically, it is **not a new AI model** — it uses the same Claude (Opus 4.8) everyone already has access to. What's new is the *harness*: the scaffolding that turns a general-purpose model into a purpose-built scientific tool. --- ## What changed — the key features | Feature | Detail | |---|---| | **60+ integrated scientific databases** | Pre-wired access to genomics, proteomics, structural biology, and chemical databases | | **Prebuilt skills** | Specialized abilities for genomics, proteomics, cheminformatics, structural biology | | **Nvidia BioNeMo integration** | Ships with Evo 2, Boltz-2, and OpenFold3 models for biomolecular tasks | | **Local-first architecture** | Runs on macOS/Linux; connects via SSH or HPC clusters — **sensitive data never leaves the lab** | | **Verification agent** | Automatically checks citations and calculations | | **Reproducible output** | Every figure includes full history: code, environment, and conversation that generated it | | **Scalable compute** | Scales from a single laptop GPU to hundreds across clusters | | **Custom skills** | Researchers can save their own pipelines as reusable skills | | **Pricing** | Available in beta for Pro, Max, Team, and Enterprise tiers; up to **$30,000 in credits** for up to 50 research projects (applications open through July 15) | --- ## Why it matters ### 1. Analogous to Claude Code's impact on software — but for all of science Six months after Claude Code transformed how engineers write code, Anthropic is betting the same playbook works for laboratory research. As **TechCrunch** put it: "bets on workflow, not a new model, to win over scientists." ### 2. A real-world demo that stunned observers At the launch event, from a single sentence, Claude Science planned and executed a drug-discovery campaign: screened 2,200 compounds across 80 GPUs for a phenylketonuria-targeting molecule, narrowed to four candidates, and produced a go/no-go memo. Then it ran the same triage across **100 rare diseases at once**. The presenter asked: why stop at 100 when the same machinery could run 10,000? ### 3. Hands-on proof: a $26 experiment that rewrote a field's ontology **Forbes contributor John Drake** (a professor at the University of Georgia) tested Claude Science on his own research library of 490 papers on zoonotic spillover. The tool discovered the field's working vocabulary is **four times richer** than its formal ontologies — 864 of 915 relationships used in the literature had no formal counterpart. **Total cost: ~$26.** He called it "a substantial" harness where "the value is curation, and curation is what turns raw capability into science." ### 4. Pharmaceutical industry upheaval Per **STAT News** (Matthew Herper & Brittany Trang), Anthropic CEO **Dario Amodei** said biology "is going to be a general purpose technology that helps us to make sense of that complexity." STAT also reported that Anthropic simultaneously announced it will **begin developing drugs of its own** — signaling the company is both selling tools to pharma and becoming a competitor. ### 5. Privacy as a differentiator Unlike cloud-only rivals, Claude Science's local-first architecture means proprietary data (protein structures, patient data, proprietary assays) never leaves the lab's infrastructure. Only the specific context Claude needs is sent to the model. ### 6. The competitive landscape **MIT Technology Review** calls it "Anthropic's newest flagship product." It goes head-to-head with **OpenAI** (which has its own scientific tools) and **Google DeepMind** (AlphaFold ecosystem). The bet is that a superior *harness* — not a superior model — wins the scientific market. --- ## Strongest sources (ranked) | Source | Why it's strong | |---|---| | **🥇 Forbes — John Drake** | First-hand academic user. A professor literally paid $26 to run his own research through the tool and published the results. This is the most concrete evidence of real-world scientific output. | | **🥇 MIT Technology Review** | Flagship tech publication; calls Claude Science "Anthropic's newest flagship product" — authoritative framing. | | **🥇 STAT News** | Two dedicated science/health reporters (Herper & Trang) covered the pharma angle in depth, plus broke the news that Anthropic is developing its own drugs. | | **🥇 TechCrunch** | Strong analysis of the "workflow, not model" strategy — differentiates Claude Science from rivals. | | **🥇 The Decoder** | Most technically detailed breakdown of the architecture (BioNeMo, local-first, HPC, 60+ databases). | | **🥇 Reuters** | Traditional wire service — confirms the story as newsworthy at global scale. | | **🥇 Anthropic (official)** | Primary source for product specs, pricing, and availability. | --- ## Bottom line **Claude Science is not about a smarter AI — it's about wiring the AI we already have into the infrastructure of science properly.** By integrating 60+ databases, supporting local/HPC deployment, and adding scientific-specific skills (genomics, proteomics, cheminformatics), Anthropic has built what may be the most comprehensive AI workbench for researchers ever released. The early hands-on evidence (a $26 academic paper analysis, a 100-disease drug screen) suggests the bet might pay off. Whether it truly accelerates scientific discovery at scale — or just automates busywork — will be determined by how labs actually adopt it in the coming months.