With 27 billion parameters, the Faraday agent reproduces scientific results better than Claude Opus 4.8 and GPT-5.5

Inherent, founded by four former Google DeepMind employees, relies on reinforcement learning rather than size to also assess the "taste" of a research agent.

Inherent, a start-up created by four Google DeepMind defectors — Edward Hughes, who is also chief scientist, Louis Kirsch, Kaloyan Aleksiev and Tantum Collins — says its agent Faraday is ahead of much heavier systems on a demanding task: autonomously reproducing the results of scientific papers. The underlying model is a Qwen 3.6 with 27 billion parameters, an order of magnitude below Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5, both of which the company says are outdated.

The thesis defended is that the reproduction of an experimental result cannot be reduced to encyclopedic knowledge: it requires knowing which experiments are worth conducting. Inherent claims to explicitly evaluate this "research taste", i.e. the instinct of selecting relevant leads, and obtains it by reinforcement learning rather than by handwritten rules.

If the result is outside the conditions chosen by the company, it fuels an open debate: an agent's ability to conduct research would depend more on the training of its decision policy than on the gross volume of parameters. The company had announced a $50 million seed round in May 2026.