London AI lab Inherent, founded by Google DeepMind alumni, has announced that its newly released AI agent, Faraday, has reportedly outperformed larger models from Anthropic and OpenAI. Faraday's reported success lies in its ability to independently reproduce the findings of published scientific papers without prior knowledge of the answer.
This achievement was reportedly accomplished using a comparatively small model, Qwen 3.6, which has 27 billion parameters. This is significantly smaller than Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5, which are described as much larger, frontier-scale systems.
Edward Hughes, cofounder and chief scientist, stated that while beating other AI systems was not the primary goal, the method of building Faraday was of particular interest. The company's broader objective is to develop AI capable of discovering new scientific knowledge, with paper replication serving as a foundational training exercise for human scientists.
Inherent, which recently emerged from stealth with a $50 million seed round, employs reinforcement learning to train its AI systems. This method rewards good outcomes rather than providing explicit rules, aiming to develop agents with "research taste" – an instinct for designing and selecting worthwhile experiments.