SandboxAQ Just Made Drug Discovery Accessible to Anyone—No PhD in Computing Required

SandboxAQ was founded roughly five years ago as a spinout from Alphabet, the parent company of Google. Eric Schmidt, the former CEO of Alphabet, serves as the company's chairman. Its advisory board includes Ray Dalio, Marc Benioff, and Yann LeCun. Since its founding, SandboxAQ has raised more than $950 million, including a $450 million Series E round backed by T. Rowe Price, Google, IQT, Paladin Capital, BNP Paribas, Breyer Capital, and S32.

The company describes its focus as the $50 trillion quantitative economy: industries where decisions depend on physics, chemistry, and mathematics rather than language or pattern recognition. Its models are already deployed at major pharmaceutical and industrial companies, with active programs covering battery chemistry, catalysts, and alloys.

SandboxAQ's Large Quantitative Models are built differently from large language models. Where LLMs learn from text, LQMs are trained on physics-grounded data: high-fidelity simulations, quantum chemistry calculations, molecular dynamics, microkinetics modeling, and real-world lab experimental data. They understand how molecules behave, not how people describe them.

SandboxAQ developed its frontier models in collaboration with NVIDIA, producing the AQAffinity and AQCat model families. One data point illustrates the stakes: catalysts underpin more than 90% of commercially produced chemical products, yet designing better ones has historically required years of work by specialized scientists. The LQM approach compresses that timeline by running simulations that would otherwise require custom computational infrastructure and deep domain expertise to configure.

The first SandboxAQ model accessible through Claude is AQCat Adsorption Spin, focused on catalyst discovery. Two additional models are in development: AQPotency and AQCell, which together form a drug discovery suite built for biopharma researchers.

Researchers can join a waitlist for access. Once enrolled, they interact with SandboxAQ's simulations by describing their scientific question to Claude in plain language, without writing specialized code or configuring external computing infrastructure. The integration does not require a PhD in computing to operate. That is precisely the point: the barrier to running these simulations dropped from 'requires a specialist' to 'requires a question.'

SandboxAQ's models already work. Active programs are underway at major pharmaceutical companies, with demonstrated advances in battery chemistry, catalysts, and alloys. The limitation was not performance; it was interface. Their models required specialized scientists and the ability to write complex code.

Nadia Harhen, SandboxAQ's General Manager of AI Simulation, described the typical customer: researchers who have already tried every conventional software option and found it insufficient. The target users are computational scientists, research scientists, and experimentalists at large pharmaceutical and industrial companies. These are people who understand the science deeply but have been blocked by the complexity of the computing layer.

The integration targets researchers across biopharma, advanced materials, energy, and financial services. In drug discovery, the numbers are stark: finding a single viable drug molecule typically takes close to a decade and costs billions of dollars, and most candidate molecules fail during clinical development. SandboxAQ's LQMs address the early-stage simulation problem, where researchers need to predict molecular behavior before committing to expensive lab work.

Beyond pharma, the models cover battery chemistry, catalysts, alloys, and materials design. SandboxAQ's broader business also spans cybersecurity, navigation, and financial risk modeling, though the Claude integration focuses first on simulation-heavy scientific research where the LQM advantage over conventional software is most pronounced.

Jack D. Hidary, CEO of SandboxAQ, framed the integration plainly: 'Now, researchers can access frontier physics-based models directly.' Partha P. Mukherjee, a professor at Purdue University who works with simulation tools in applied research, noted that the integration 'removes one of the key barriers' to running advanced simulations within scientific workflows. Both observations point to the same structural shift: the constraint was always the interface, not the underlying science.

This move signals how Anthropic is betting on the future: agents need reach. Models are only useful if they can connect to real tools and real data. MCP was built for exactly this purpose, and SandboxAQ's integration is one of the most practically impactful applications of MCP released to date. It turns academic-grade scientific simulation into a tool that any qualified researcher can operate.

For business leaders in life sciences, materials, or energy, this is a concrete illustration of the ROI case for building on Claude as a platform. The model's reach expands with every credible external integration. The cost of running complex scientific simulations drops with every integration that eliminates the need for specialized infrastructure. If you work in biotech, pharma, or materials science, Claude just became a significantly more capable research tool.