A medicinal chemist who has never written a line of code can now ask a physics-based simulation how a drug candidate binds its target — in plain English. SandboxAQ, the Palo Alto company that spun out of Alphabet, said Monday it has integrated its physics-based AI models with Anthropic’s Claude, letting researchers run drug discovery and materials science workloads through conversational prompts. The company’s argument is explicit: in AI-driven science, the bottleneck is not model quality but access.
Until now, running the advanced models for new drug and materials discovery required specialized scientists and the ability to write complex code. With Claude serving as a natural-language interface to SandboxAQ’s platform through the Model Context Protocol, any researcher can reach the same capability without infrastructure setup, moving from hypothesis to computational answer in ordinary language.
The move comes at a moment when drug discovery is among the most expensive pursuits in modern industry: finding a single viable molecule can take a decade and cost billions, and most candidates still fail. A generation of AI startups has promised to shorten that timeline, and most have built tools that help researchers who already know how to use them. SandboxAQ’s bet is that the interface, not the model, has been blocking wider adoption.
SandboxAQ builds what it calls Large Quantitative Models — AI systems trained on real-world lab data and scientific equations rather than text patterns, engineered for what the company sizes as a more than $50 trillion “quantitative economy” spanning biopharma, financial services, energy and advanced materials. The distinction matters technically: LLMs predict patterns in text, while LQMs are designed to respect the rules of the physical world. That makes them suited to tasks like predicting how a catalyst behaves or whether a molecule binds a target, where pattern matching alone produces confident but wrong answers.
Several frontier models are already in the pipeline. AQAffinity and AQCat, both developed in collaboration with Nvidia, support drug binding and catalysis work, and AQCat Adsorption Spin is live. A suite of drug discovery models is coming next: AQPotency, which screens thousands of candidate compounds computationally to prioritize the most promising ones, and AQCell, which simulates how living cells respond to drug candidates, predicting whether a compound activates the right biological pathway and flagging potential liver toxicity. SandboxAQ is also running active programs at major pharmaceutical companies, with demonstrated advances in battery chemistry, catalysts and alloys.
The move lands in a crowded field. Venture-backed companies such as Chai Discovery and Isomorphic Labs have raced to build better protein and molecule models, and SandboxAQ’s counter-argument is that model quality is no longer the binding constraint — reach is. Nadia Harhen, the company’s general manager of AI simulation, said the integration lets pharma and biotech users “run workflows that previously required weeks of computational setup in hours,” regardless of technical background. Customers, according to Harhen, come to SandboxAQ “because they’ve tried all the other software out there, and the complexity of their problem is such that it didn’t work” — problems that need physics-grounded computation rather than statistical approximation.
The pitch to pharma is straightforward. Drug companies already run hundreds of simulations in the hunt for a candidate; the question is how many scientists can commission them. In the company’s telling, an organization with a thousand bench scientists but a handful of computational specialists can now put a physics-grounded simulator in front of most of them — and the waitlist will test whether that changes which programs actually get run.
Outside voices backed the framing. Partha Mukherjee, a Purdue professor who directs the Center for Advances in Resilient Energy Storage, said the integration “removes one of the key barriers between a researcher’s scientific intuition and rigorous physics-grounded computation.” Woody Sherman, chief innovation officer at PsiThera and executive committee chair of the OpenFold Consortium, called it “a critical barrier between researchers and the frontier of computational science.” Chief Executive Jack Hidary said the models bring “the rigor of first-principles quantum chemistry to a conversational interface.”
SandboxAQ was established as an Alphabet spinout roughly five years ago and is chaired by Eric Schmidt, Google’s former chief executive. It has raised more than $950 million from investors, including a $450 million Series E, and has built business lines beyond science: an AI-native cybersecurity unit, medical research tools and navigation systems, with financial services and risk modeling modules set to go live soon.
The integration is a distribution play as much as a technology one. Anthropic’s platform has become a common entry point for enterprise AI, and SandboxAQ is betting that putting its models inside an interface scientists already use will pull in a broader R&D audience than selling specialized software to computational chemists. Access is free to start — the company is running a waitlist for users who want the models in Claude — and the commercial model will be tested by whether researchers who start with prompts graduate to paying workloads. If the bet holds, the next wave of AI drug discovery will not be measured by benchmark scores but by how many bench scientists can ask a question in plain English and get a physics-grounded answer.
