If AI drug discovery models are already powerful, why does so much of their value still sit behind specialist infrastructure, code, and computational workflows?
That is the real question behind SandboxAQ’s tie-up with Anthropic. The company is integrating its Large Quantitative Models, or LQMs, with Claude, giving researchers a natural-language route into scientific models used for drug discovery and materials science, according to TechCrunch. The near-term product detail matters: SandboxAQ says AQCat Adsorption Spin is accessible via Claude now, while drug discovery models including AQPotency and AQCell are coming soon.
Is SandboxAQ Competing on Better Models, or Better Access to Them?
SandboxAQ is not simply pitching another AI model for pharma. It is making a distribution argument.
The company’s bet is that the limiting factor in AI-enabled drug discovery may not be raw model capability alone. It may be whether chemists, biologists, computational scientists, and experimental teams can actually use those models inside normal research workflows.
That puts SandboxAQ in a different lane from companies such as Chai Discovery and Isomorphic Labs, which TechCrunch describes as “well-funded bets on better models.” SandboxAQ is still building scientific models, but the Claude integration shifts the emphasis toward access.
“For the first time, we have a frontier [quantitative] model on a frontier LLM that someone can access in natural language,” Nadia Harhen, SandboxAQ’s general manager of AI simulation, told TechCrunch.
The nuance: this is not “Claude discovers drugs.” Claude becomes the interface layer. SandboxAQ’s LQMs remain the scientific engine.
That distinction matters because enterprise adoption often breaks on usability before it breaks on theory. Previously, SandboxAQ users had to bring their own digital infrastructure to run the models. Claude gives the company a way to put complex simulation tools behind plain-English prompts.
For more on how Claude is becoming a strategic surface for technical work, see MLXIO’s coverage of Anthropic Grabs Andrej Karpathy for Claude AI Race and the risks around over-reliance in Claude Code Exposes the New Coding Risk: Blind Trust.
How Much Scientific Work Can Claude Actually Translate?
Claude’s role is strongest where scientific intent needs to become technical execution.
SandboxAQ’s LQMs are described as “physics-grounded” models. That means they are built around scientific equations and lab data rather than text-pattern prediction alone. The company says the models can run quantum chemistry calculations, simulate molecular dynamics, and model microkinetics, the molecular-level study of how chemical reactions unfold.
In practice, Claude could reduce the friction between a researcher asking a question and the model running the relevant workflow. A user might not need to write complex code or manage bespoke infrastructure to test a hypothesis, compare candidates, or interpret outputs.
SandboxAQ’s own announcement frames this as a move from computational setup to usable scientific interaction:
“Now, researchers can access frontier physics-based models directly inside the AI tools they already use, with no additional infrastructure, code or barriers,” said Jack D. Hidary, CEO of SandboxAQ.
MLXIO analysis: the interface is not a cosmetic layer if it changes who can operate the system. A tool used only by computational specialists has one adoption curve. A tool that experimentalists and research scientists can query directly has another.
Still, the Claude layer does not remove the hard part: proving that outputs hold up when translated into real-world lab work.
Why Does the Cost Structure Make Access a Serious Business Question?
Drug discovery remains a brutal filter. TechCrunch notes that finding a single viable molecule can take a decade and cost billions, while most candidates still fail.
That is why even narrow improvements can matter. Better hit identification, lead prioritization, toxicity flagging, or candidate triage can change how teams allocate lab time and modeling resources. The value is not only in finding a molecule faster. It is also in killing weak candidates earlier.
SandboxAQ is positioning LQMs for what it calls the “quantitative economy”:
“Trained on real-world lab data and scientific equations, LQMs are AI models engineered for the quantitative economy, a $50+ trillion sector spanning biopharma, financial services, energy, and advanced materials,” the company said.
The company has raised more than $950 million, spun out of Alphabet roughly five years ago, and counts Eric Schmidt as chairman. That capital base supports a broad company rather than a single-purpose drug discovery startup: SandboxAQ also has cybersecurity, medical research, navigation, and soon financial services and risk modeling modules, according to its announcement.
The business logic is clear. If Claude lowers the access barrier, SandboxAQ can sell scientific computing to more users inside the same customer organization.
Are SandboxAQ, Chai Discovery, and Isomorphic Labs Fighting the Same Battle?
Not exactly.
The competitive divide is becoming clearer: one camp races to prove frontier scientific model performance; another tries to embed models where enterprise teams already work. SandboxAQ is trying to do both, but the Claude deal highlights the second path.
| Company | Source-supported positioning | Strategic emphasis |
|---|---|---|
| SandboxAQ | Builds proprietary LQMs for drug discovery, materials science, cybersecurity, navigation, and other sectors | Access, enterprise workflows, physics-grounded simulation |
| Chai Discovery | Named by TechCrunch as a well-funded company focused on better models | Model capability in AI biology and drug discovery |
| Isomorphic Labs | Named by TechCrunch as another well-funded bet on better models | Scientific model performance and drug discovery workflows |
SandboxAQ’s differentiator is its claim that LQMs are built from the ground up with physics-grounded training data, including high-fidelity simulations and lab data augmentation. The company says it owns the models outright and connects them into automated workflows.
That may appeal to large pharmaceutical and industrial customers dealing with problems that are too complex for off-the-shelf tools. Harhen told TechCrunch:
“Our customers come to us because they’ve tried all the other software out there, and the complexity of their problem is such that it didn’t work or didn’t yield positive results for them when that translation went to take place in the real world.”
MLXIO analysis: this is a classic enterprise AI wedge. Don’t just claim better intelligence. Claim better translation from model output to operational decision.
Why Start With Catalysts Before Claude-Enabled Drug Discovery?
The current Claude-accessible model is AQCat Adsorption Spin, not the full drug discovery suite.
That matters. AQCat focuses on adsorption energy calculation, which measures how strongly molecules bind to a catalyst surface. SandboxAQ calls it the critical first step in a catalyst discovery workflow because it helps researchers prioritize candidates before committing more expensive modeling and lab resources.
The company says catalysts underpin more than 90% of all commercially produced chemical products, with potential relevance to green hydrogen, sustainable aviation fuel, fertilizer production, plastics recycling, and more.
Drug discovery comes next. SandboxAQ says AQPotency will help users computationally identify and prioritize promising drug candidates, while AQCell will simulate how living cells respond to drug candidates across thousands of compounds, including pathway activation and potential liver toxicity.
That sequencing is important. The Claude integration is being proven first through a materials-science workflow, then extended into pharma.
What Will Pharma Teams and Investors Be Watching Before They Believe It?
For pharma R&D teams, the pitch is productivity. The questions will be reliability, explainability, data governance, intellectual property protection, and whether Claude-mediated workflows produce outputs that can be defended inside serious research programs.
For AI drug discovery startups, the message is sharper: model architecture is not the only moat. Interface, distribution, and workflow embedding may decide which tools get used beyond the computational science group.
For investors, SandboxAQ gains a broader narrative. It is not only an AI biology company or a quantum-inspired simulation shop. It is trying to turn physics-grounded models into enterprise software surfaces through Claude.
Anthropic also gets a more specialized enterprise use case. Claude is not just a general assistant in this setup. It becomes a front end for domain-specific scientific computation. That is a different kind of value proposition than generic productivity.
The evidence to watch is practical. Do researchers on the waitlist move from question to defensible answer faster? Do pharma teams use AQPotency and AQCell when they arrive? Do Claude-based workflows expand usage beyond computational specialists?
If SandboxAQ is right, the next phase of AI drug discovery will not be won only by the company with the flashiest benchmark. It will be won by the company that makes complex scientific models usable at scale — without pretending that easier access is the same thing as validated science.
The Bottom Line
- SandboxAQ is trying to make advanced drug discovery and materials science models usable without specialist computational workflows.
- The Claude integration could broaden access for chemists, biologists, and experimental teams working in normal research environments.
- The move highlights a key AI adoption challenge: usability may matter as much as model performance in enterprise science.
