Inherent's Faraday AI Beats OpenAI and Anthropic

London-based AI lab Inherent has revealed that its new AI agent, Faraday, successfully outperformed frontier models from OpenAI and Anthropic at replicating published scientific research. Despite running on Qwen 3.6, a comparatively small model with just 27 billion parameters, Faraday achieved higher accuracy in independently reproducing paper findings than OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8. The achievement comes shortly after the startup, founded by Google DeepMind alumni, emerged from stealth with a $50 million seed round.

To achieve this, Inherent focused on reinforcement learning to instill what chief scientist Edward Hughes calls research taste, or an intuitive grasp of which experiments are worth designing. Rather than building every tool from scratch, Faraday utilized OpenAI's GPT-5.5 Codex to handle coding tasks, mimicking how human scientists rely on existing software. This reward-based training approach is designed to help the agent generalize its skills across multiple scientific disciplines, moving toward Inherent's ultimate goal of discovering entirely new scientific knowledge.

For AI practitioners and researchers, Faraday's success demonstrates that massive parameter scale is not the only path to high-level reasoning. By combining a highly optimized 27-billion-parameter base model with targeted reinforcement learning and external tools, developers can build highly capable, domain-specific agents at a fraction of the computational cost of frontier systems. This shift could democratize scientific AI assistant development, allowing smaller labs to run sophisticated research workflows without massive infrastructure budgets.

Inherent currently operates with a team of 12 employees in London's King's Cross district, with plans to expand its headcount to between 20 and 25 by the end of the year. The startup's founders, including Hughes, Louis Kirsch, Kaloyan Aleksiev, and Tantum Collins, hope to leverage London's dense talent pool to continue developing world models, even as they navigate local hiring hurdles like the UK's common practice of garden leave.