Faraday d'Inherent Surpasses OpenAI and Anthropic in Scientific Research

The British company Inherent, founded by former DeepMind employees, is focusing on a "teammate" agent trained through reinforcement learning for scientific research. The company claims that Faraday has outperformed models from Anthropic and OpenAI in reproducing results, all while operating with a model of 27 billion parameters. Based in London, it plans to quickly expand a team that works entirely on-site.

Inherent is hiring in London and challenging "garden leave"

Inherent currently has twelve employees who all work on-site in King’s Cross, and Edward Hughes believes London is the place to be for AI. He is optimistic about the local talent density, and the company aims to grow to about 20 to 25 people by the end of the year. Hughes advocates for the abandonment of the British "garden leave," a practice that can prevent a resigning employee from joining or starting a competitor for months. He points out that American researchers generally face this issue less, giving startups in the U.S. an advantage in recruiting profiles that have left their positions. He clarifies that this position is personal, as he himself was affected before finding a way around this constraint and co-founding Inherent with two other former DeepMind employees and a fourth co-founder. The company also has ambitions in global models. In a context where Demis Hassabis has taken on a new role that may leave some DeepMind employees uncertain, Inherent's hiring campaign could make it a landing spot for those considering a change.

The "teammate" agent is trained with a taste for research through reinforcement learning

Inherent claims to train its agents primarily through rewards rather than direct study of scientific practice, betting on better generalization towards contributions across multiple disciplines. Teaching a "taste" for research is presented as challenging, and reinforcement learning—which rewards results rather than imposing rules—serves as the methodological foundation. Edward Hughes aims for an AI scientific agent capable of judging which experiments to launch and how to design them, and this focus also guides what the company does not build. Faraday relies on OpenAI's Codex GPT-5.5 coding tool rather than a proprietary tool, an approach the company likens to the practice of scientists using existing software. Inherent also seeks to avoid agents that comfort the user, drawing inspiration from a teammate who conducts experiments independently and then returns to confront the results.

Claimed benchmark: Faraday surpasses Claude Opus 4.8 and GPT-5.5

Inherent claims that Faraday has recently surpassed larger systems from Anthropic and OpenAI on a specific task: independently reproducing published results without knowing the answer in advance. The comparison explicitly cites Claude Opus 4.8 and GPT-5.5. The agent operates on Qwen 3.6, a model described as relatively small with 27 billion parameters, while the number of parameters is a common indicator of size and often training costs. The success criterion went beyond accuracy: the team wanted to observe a "taste for research," meaning an instinct about which experiments to conduct and their design. Edward Hughes reminds us that reproducing articles is a standard exercise among scientists, with many PhD students starting there, and clarifies that the goal was not to outperform other AIs but to understand the construction method.

Funding and scientific leadership

Inherent emerged from discretion a few weeks ago by announcing a $50 million funding round. The lab, founded in London by former Google DeepMind employees, has Edward Hughes as its scientific director and co-founder.