Anthropic's Claude Science, launched in late June 2026, promises to accelerate research workflows by integrating AI directly into genomics, proteomics, and structural biology tasks. Early users report significant speed gains in analyzing genetic datasets and assembling literature reviews, but they also highlight a fundamental gap: the AI still struggles to think critically about data the way trained scientists do.

What Is Claude Science and How Are Researchers Using It?

Claude Science is a specialized toolkit built on Anthropic's Claude AI model, designed specifically for laboratory research. The tool comes with more than 60 pre-configured connectors that handle workflows across genomics, proteomics, and structural biology. It can analyze and produce code, generate images through integration with BioRender, a popular figure-creation tool, and suggest next steps for wet lab experiments.

Stephen Francis, a neuroscientist at the University of California, San Francisco, received early access to Claude Science and has been using it to analyze massive genetics datasets related to glioma, a type of brain cancer. Francis explained that the tool excels at scraping through obscure databases to find supporting evidence for genetic findings and then presenting a ranked list of possible next analysis steps with explanations for each recommendation.

"Very often it'll get my top three that I'm thinking of, and it'll add a couple more on that I didn't think of," said Francis.

Stephen Francis, Neuroscientist at University of California, San Francisco

Jerome Lecoq, an associate investigator with the Allen Institute in Seattle, has used Claude Science primarily to create literature reviews. He noted that the AI assembles comprehensive reviews far faster than humans could, while also documenting exactly how it derived each fact and which prompts led to the final output.

Where Does Claude Science Excel, and Where Does It Fall Short?

The toolkit offers several advantages over the standard Claude interface. Kelvin Lau, a researcher at the Swiss Federal Technology Institute of Lausanne, emphasized that Claude Science shows its reasoning step-by-step, explaining which code it uses and why. He also noted that the tool appears to run more analysis locally on a user's device, a feature that appeals to privacy-conscious researchers, even if it means slightly longer computation times.

Anthropic has made specific adaptations to how Claude Science generates and presents information. Early testers requested a "replication" button that would review and check the AI's own output with a skeptical eye, and Anthropic implemented this feature, leading to noticeable accuracy improvements.

However, significant limitations remain. Lecoq pointed out that while Claude Science can assess how many papers support or reject an idea, it struggles to evaluate the quality of individual papers, a core skill in human-led literature reviews. He also noted that the AI cannot access paywalled academic papers, a persistent tension between AI developers and academic publishers.

Francis reported that Claude Science refused to respond to some of his team's prompts about herpesviruses and brain tumors, citing biosecurity protocols. While he understood Anthropic's caution, he argued that the guardrails are currently too restrictive for legitimate research.

How to Improve AI's Scientific Reasoning Capabilities

• Incorporate Scientific Methodology into Training: Lecoq emphasized that reinforcement learning data used to train Claude and similar models must incorporate scientific methodology, teaching the AI to question data rather than simply predict the next word.
• Develop Better Quality Assessment: AI systems need to learn how to evaluate the quality and rigor of research papers, not just count how many papers support a hypothesis.
• Refine Biosecurity Guardrails: Researchers suggest that Anthropic and other AI companies should calibrate safety protocols more carefully to distinguish between legitimate research and actual biosecurity risks.

Lecoq captured the core challenge: "It's not about predicting the next token anymore. It's about doubting the next token." This distinction highlights why AI, despite its speed advantages, has not yet fully replicated the skeptical, questioning mindset that defines scientific training.

Lecoq

What Do Researchers Say About the Future of AI in Science?

All three early-access researchers said they plan to continue using Claude Science, recognizing its productivity benefits. However, they also raised concerns about how widespread AI adoption might reshape the research landscape. Lau posed a critical question for academic institutions: "Is your mission more productivity or to teach the next generation?"

There is concern that grant-stressed lab heads might view AI tools as a shortcut to faster workflows, potentially reducing hiring of new scientists. This raises broader questions about the future structure of research teams and whether institutions will prioritize speed over mentorship and training.

Despite these reservations, the tool represents a significant step toward integrating AI into everyday research practice. The speed gains are real, the transparency improvements are measurable, and the ability to suggest novel analysis pathways has genuine value. What remains uncertain is whether AI can be trained to embody the critical, questioning spirit that defines rigorous science, or whether it will remain a powerful but ultimately uncritical assistant.