Anthropic Launches Claude Science: An AI Workbench for Researchers, Claiming 10x Speedup in Tests

On June 30, 2026, Anthropic released Claude Science, an AI workbench designed for scientists—not a new model. It integrates literature search, data analysis, chart generation, manuscript writing, and other research workflows into a single environment, supporting macOS and Linux. Pro, Max, Team, and Enterprise users can access it immediately.

Core Features and Workflow

• Reproducibility: Each chart comes with the generating code, runtime environment, and conversation history. Supports natural language modifications (e.g., "remove gridlines"), and results are publication-ready.
• Intelligent Compute Scheduling: Automatically manages local, SSH remote, or HPC cluster resources, scaling from a single GPU to hundreds of GPUs. Sensitive data never leaves the original system.
• Out-of-the-Box: Pre-installed with 60+ scientific skills and connectors, directly linking to hundreds of databases such as UniProt, PDB, Ensembl, ClinVar, ChEMBL, and GEO. Integrates NVIDIA BioNeMo Agent Toolkit, supporting specialized models like Evo 2, Boltz-2, and OpenFold3.
• Review Mechanism: Built-in review agent dedicated to checking citation accuracy and computational correctness. All critical operations require user-by-user authorization, ensuring human-in-the-loop.

Test Performance

• In a brain-computer interface literature review task, Claude Science first formulated a six-step plan, then automatically fetched metadata from OpenAlex after confirmation. When rate-limited, it switched to Crossref. Ultimately, it filtered 1,218 relevant papers (62% with abstracts), performed clustering and trend analysis, and output a Chinese report, charts, and CSV files.
• The entire process is visible and traceable, proactively noting data source limitations and suggesting supplementary authoritative databases.

Early User Feedback

• Allen Institute: Neuroscientist Jérôme Lecoq's team compressed long-form review writing from nearly 2 years to a few weeks, using an "actor-critic" dual-agent model (one writes, one critiques), producing multiple hundred-page results.
• UCSF Brain Tumor Center: Glioma molecular epidemiology analysis time reduced to 1/10 of the original, with independently verified reliable results.
• Manifold Bio: Used for tissue-targeted drug target screening, evaluating surface expression, transport, and safety, outperforming general coding assistants in efficiency.

Industry Comparison and Cautious Voices

• OpenAI: Launched GPT-Rosalind (proprietary fine-tuned model, enterprise-only) and GeneBench-Pro benchmark to test model scientific judgment.
• Google: Holds exclusive models like AlphaFold and AlphaGenome; Gemini for Science integrates 30+ scientific databases.
• Anthropic: Adopts a subscription-based, open, and workflow-first route, while launching a drug discovery initiative funding 50 AI for Science projects (up to $30,000 each).
• Northeastern University's Jared Auclair cautions: AI is still a "co-pilot"; it may hallucinate or miss details when interpreting regulatory guidelines or designing experiments, so careful use is advised.