Inherent, a London-based AI laboratory founded by Google DeepMind alumni, unveiled Faraday, an AI agent built on a 27 billion parameter model that outperformed larger systems from Anthropic and OpenAI in scientific research replication tasks.
Faraday independently reproduced findings from published scientific papers without prior knowledge of expected results. The startup emerged from stealth weeks ago with a $50 million seed round.
The agent runs on a Qwen 3.6 model—a significantly smaller architecture than frontier-scale systems like Anthropic's Claude Opus and OpenAI's GPT-5.5, against which Inherent measured performance. Because model parameters correlate directly with training costs, the achievement signals a path toward competitive performance with lower computational spend.
"What was most interesting to us about this was not so much the result of beating those frontier agents—which of course we liked—but actually the way we went about building this," said Edward Hughes, cofounder and chief scientist at Inherent.
Inherent's strategy extends beyond replicating existing work. The company aims to build AI capable of discovering new scientific knowledge, with paper replication serving as foundational training similar to PhD research.
Faraday's training uses reinforcement learning, which rewards successful outcomes rather than applying explicit rules. This approach is designed to instill "research taste"—the ability to identify valuable experiments and design them effectively.
Inherent demonstrates a capital-efficient approach to product development: rather than building proprietary coding tools, Faraday leverages OpenAI's GPT-5.5 Codex, mirroring how human scientists use existing software rather than building every component.
