27B Ex-DeepMind Agent Outperforms Claude Opus & GPT-5.5 in Science Questions.

UK-based startup Inherent, which was founded by a team from Google DeepMind, has secured $50M in seed capital and launched Faraday, an AI agent that acts as a scientist with 27 billion parameters. The model is based on Qwen 3.6 and trained using reinforcement learning (trial & error) to learn how to conduct scientific experiments.

Faraday is a startup founded by ex-DeepMind researchers Edward Hughes (chief scientist), Louis Kirsch, Kaloyan Alexiev, and Tantum Collins. In experiments that involved replicating work from published scientific studies, Faraday beat models such as Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5, which are much larger than it.

If this research can be replicated, it dramatically changes the dynamics of the AI for science competition. Turns out domain specificity is more important than sheer scale. Faraday is being developed as an “AI research assistant” — the agent autonomously pursues and shares results on questions.