Anthropic launched Claude Science yesterday, and the early reviews from actual working scientists are encouraging. One researcher who tested it said Claude "became Ph.D.-level in some areas of physics, chemistry and biology." That's not a marketing line from Anthropic, that's a professor talking about a tool he just used.
The launch landed June 30, 2026, and the pitch is straightforward: take the sprawling mess of tools and databases a scientist has to juggle and put it in one place. Claude Science wires together more than 60 scientific databases and computation tools into a single workspace, so instead of bouncing between a dozen systems, a researcher works in one. It drops Anthropic directly into a fight with OpenAI, whose GPT-Rosalind science model came out back in April 2026.
What it actually does
The use cases early testers point to are the heavy stuff, not toy demos. Drug discovery and drug repurposing. Protein research tied to ALS, the disease also known as Lou Gehrig's. Identifying organic pollutants. Chewing through mass spectra data. Cancer biology. These are real bench-science problems where the bottleneck isn't writing code, it's reasoning across huge piles of specialized data.
The beta is only available for Mac and Linux right now, according to Northeastern's Zhenyu Tian. So if you're on Windows, you're waiting.
The scientists testing it are impressed
The reactions gathered by Northeastern University, where a group of researchers got early hands-on time, run genuinely enthusiastic. Jeffrey Agar, who says "I use AI a lot, sometimes for many hours a day," is the one who called Claude "Ph.D.-level in some areas of physics, chemistry and biology." Coming from someone who lives in AI daily, that is a strong claim.
Michael Pollastri, also at Northeastern, didn't hedge much either: "Claude Science looks like it's going to be an unbelievable tool," and he said it could increase the pace of experiments "by orders of magnitude." Tian pointed at a specific pain point it could crack, the identification of small molecules, which is one of those quietly enormous bottlenecks that slows a lot of work down.
The claim that matters: years off drug development
Jared Auclair, another Northeastern researcher, floated that AI tools like this could cut drug development from the current 10 to 15 years on average down to two to five years. If that holds up even partway, it isn't an efficiency tweak, it's a different timeline for getting medicine to people. That is the actual promise sitting underneath all the workspace-and-database talk.
Now… hitting the brakes
The same people who are excited are also careful. Auclair paired his optimism with a warning that AI can "hallucinate or miss nuance in regulatory guidance," a real problem when you're in an industry where a wrong answer isn't a bug, it's a failed trial or a rejection. He stated, "a co-pilot that requires a skilled pilot." The tool is powerful, but it does not remove the expert, it makes the expert faster.
Bryan Spring struck the same chord, saying AI has the potential to "significantly accelerate scientific discovery" without replacing scientists. And not everyone is diving in. Researchers including Donald O'Malley, Dong Sijia, and George O'Doherty reported minimal AI use so far, with Sijia specifically waiting on better data protection before trusting it with real work.
Why this one is a big deal
The same class of models that has been eating software development is now getting pointed at the hard sciences, and the people qualified to judge it are saying it clears a real bar. We already got a preview of this when Anthropic's Fable 5 built a genomics model that beat a published Science paper, so a dedicated science product was the obvious next move.
The thing to hold onto is that "co-pilot that requires a skilled pilot" line, because it's the same truth whether you're doing cancer biology or shipping a web app: the AI gets you there faster, it does not check your work for you. The scientists who win with Claude Science will be the ones who already know enough to catch it when it's confidently wrong, which is the exact same skill that separates the builders who last from the ones who ship a mess. New domain, same rule.
Sources
Northeastern Global News: Anthropic launches Claude Science - The June 30, 2026 launch of Claude Science integrating more than 60 scientific databases and computation tools, its positioning against OpenAI's GPT-Rosalind, the Mac and Linux beta availability, and reactions from Northeastern researchers including Jeffrey Agar's "Ph.D.-level" assessment, Michael Pollastri calling it an "unbelievable tool" that could speed experiments by orders of magnitude, Zhenyu Tian on small-molecule identification, Jared Auclair's estimate that drug development could drop from 10 to 15 years to two to five years along with his "co-pilot that requires a skilled pilot" caution about hallucination and regulatory nuance, Bryan Spring on accelerating discovery without replacing scientists, and Donald O'Malley, Dong Sijia, and George O'Doherty on limited use and data-protection concerns.
