Inherent, a British artificial intelligence laboratory founded by former DeepMind researchers, said its new agent Faraday outperformed systems from Anthropic and OpenAI on a test of replicating published scientific papers. The company described Faraday as an AI teammate built to reconstruct experiments and findings from existing research rather than generate original claims. The announcement casts the lab as a specialist in scientific workflow tools at a time when larger U.S. firms still dominate general-purpose models.
Faraday is designed to read papers, recover methods, and attempt to reproduce results, a task that has become a practical benchmark for whether language models can follow scientific procedure. Inherent said the agent’s performance exceeded that of comparable offerings from Anthropic and OpenAI, though it did not immediately release a full independent evaluation. Replication is demanding work: it requires parsing methods sections, tracking variables, and executing or simulating experimental steps without inventing data.
The company framed successful replication as a stepping stone toward tools that could ease laboratory work, not as a substitute for human scientists. If an agent can reliably reconstruct prior studies, researchers might spend less time re-deriving methods and more time on new questions. That prospect remains unproven. Inherent has not shown that Faraday can design novel experiments or resolve conflicting findings.
Replication is a narrow skill, but it is one of the few that can be checked against the public record.
Frontier labs have invested heavily in agents that browse the web, write code, and operate software, even as public documentation of how they would contain unexpected model behavior remains limited. Faraday’s narrower brief—reproducing papers rather than acting freely across the open internet—may reduce some of those risks, though any system that runs code or contacts external tools still needs oversight. Independent researchers will need access to Inherent’s evaluation suite before the outperformance claim can be weighed against Anthropic’s and OpenAI’s own scientific agents.
Inherent joins a crowded group of startups founded by DeepMind alumni, many of which have tried to compete on reliability rather than raw scale. The lab has not disclosed Faraday’s training data, compute budget, or commercial terms. For now, the product will be judged by whether working scientists can use it to reproduce papers they already trust, and whether those reproductions hold up under human review.
