Inherent introduces Faraday, an AI agent for scientific research

London-based startup Inherent, founded by former Google DeepMind members, has introduced Faraday, an artificial intelligence agent designed for scientific research. According to a preprint submitted to arXiv, Faraday demonstrated the ability to replicate scientific findings from published papers, outperforming larger models from OpenAI and Anthropic in specific tasks.

In internal tests, Faraday reportedly surpassed Claude Opus 4.8 and GPT-5.5 in 73% of machine learning tasks similar to training data and in 60% of science-applied AI tasks. Unlike massive models that rely on sheer scale, Faraday utilizes the Qwen 3.6 27B model as a base, prioritizing reasoning efficiency and autonomous decision-making through reinforcement learning.

The benchmark used by Inherent involved 310 tasks extracted from 100 scientific articles. While the results show promise for creating agents capable of discovering new knowledge, the study has not yet undergone independent validation, and the preprint does not currently provide the full code, weights, or a reproducible environment for external verification.