An AI agent from a London startup has outdone Anthropic and OpenAI systems at a core scientific chore: re-deriving the results of published papers when the answers are not supplied in advance. The team behind it left Google DeepMind to found Inherent.
The startup, which emerged from stealth weeks ago with a $50M seed round, released the agent under the name Faraday. It runs on Qwen 3.6, a model with just 27 billion parameters, while its rivals in the test were Claude Opus 4.8 and GPT-5.5, both frontier-scale. Chief scientist Edward Hughes told TechCrunch the result mattered less than the method. The team trained Faraday with reinforcement learning rather than by teaching it how science is usually done, betting that reward-based learning will generalize to its real goal: agents that design and run their own experiments.
Inherent set a higher bar than accuracy, asking Faraday to show “research taste” – an instinct for which experiments are worth running. The agent even borrowed OpenAI’s GPT-5.5 Codex for coding work, the way human researchers lean on existing tools. The company plans to grow from about a dozen employees to 20-25 by the end of the year, with hiring expected to draw from DeepMind staff unsettled by recent changes at the lab.
For now the demo doubles as a recruiting pitch. If a 27B model can keep pace with models many times its size on replication, Inherent argues, the path to AI that discovers new knowledge starts with learning to verify what is already known.
