In a move in keeping with its sustained push into various vertical industries, Anthropic on Tuesday introduced Claude Science, an AI workbench for scientific research.
Claude Science is designed to streamline scientific research by consolidating fragmented tools and databases into a single, integrated environment — one that Anthropic specifies is not an AI model in itself.
The new application is another way Anthropic is evolving its focus from mainly AI for coding. The shift in focus can be seen in Anthropic’s recent moves, first toward law and finance, then to cybersecurity and now toward applications such as science. Claude Science is not the vendor’s first foray into science. Both the Fable 5 and Mythos 5 models have capabilities in molecular biology and genomics.
“It’s encouraging that Anthropic is continuing to dedicate some of its resources to these hard problems, rather than just doubling down on the clearly economically successful tools they’re developing for AI for code,” said John Thickstun, assistant professor of computer science at Cornell University.
The AI workbench provides a unified research environment preconfigured for genomics, single-cell analysis, structural biology and cheminformatics. It also renders 3D protein structures, generates publication-ready figures and includes full code, environment details and creation history for every output. The AI workbench uses Opus 4.8 with no special access required and integrates with Nvidia’s BioNeMo Agent Toolkit and supports Model Context Protocol for custom extensions.
The emphasis Anthropic places on the workbench not being a model is important, said Chirag Shah, a professor in the Information School at the University of Washington in Seattle.
“There have been some attempts for that,” Shah said, referring to previous efforts to create new scientific models. “They haven’t panned out so much like taking a foundation model and then fine-tuning it on some domain like biology.”
He added that what Anthropic provides also isn’t completely new, as competitors such as OpenAI and Google already offer similar tools. For example, OpenAI has FrontierScience benchmark, while Google’s Gemini for Science is a suite of AI tools for scientific research.
The Aim for the Science Vertical
However, Claude Science could provide a certain level of convenience for some researchers and scientists.
“Somebody who doesn’t have all the know-how of how to create a pipeline using foundation model and different harnesses, they can now just take this off the shelf and it’s ready to go,” Shah said. “When you think about science, it comes down to creating a hypothesis, testing a hypothesis, collecting data, doing a lot of literature review, connecting the dots, data analysis. These things could be automated, or a lot of those things, and so that’s what this is,” Shah said.
Some Challenges
However, science is harder for AI to tackle than coding.
“Science is a little less readily available,” Thickstun said. He said that means vendors might not be able to access scientific information easily.
Another challenge is timing.
“AI likes to work in environments where it’s able to get feedback at a superhuman computational scale,” Thickstun added. “With a lot of scientific applications, you’re lucky to get feedback once a year.”
Moreover, science is about both the process of accessing information and its outcomes, Shah said.
“It’s not just about what comes out of it, but how you get to it,” he said.
