Inherent startup agent beat Claude and GPT at study replication with a ten times smaller model

A smaller model beats universal giants

British startup Inherent, founded by Google DeepMind researchers, showcased an agent named Faraday. Its task is not writing text, but the ability to independently replicate the results of published scientific studies without knowing the correct answer in advance. To the market's surprise, the agent achieved better results than systems built on Claude Opus 4.8 and OpenAI's GPT-5.5, even though Faraday runs on a much smaller Qwen 3.6 model with 27 billion parameters.

Training targeted research taste, not rules

The Inherent team trained the model using reinforcement learning. It didn't try to mechanically teach the agent fixed rules of the scientific method, but rewarded it for high-quality results. The emphasis was on building “research taste” (an instinct for which experiments are worth running and how they should be designed). The agent thus doesn't just repeat empty steps, but evaluates their meaning.

The agent doesn't write code itself, it delegates to Codex

Faraday doesn't operate in isolation. When it needs to write or modify code to replicate a study, it doesn't try to generate it using its primary small model. Instead, it uses OpenAI's Codex (built on GPT-5.5). It thereby mimics the behavior of human scientists, who reach for the best available software on the market instead of building their own tools.

From replication to designing new experiments

The ability to repeat an old study won't bring a revolution on its own. For startup Inherent, which plans to double its London team to 25 people by the end of the year, Faraday is merely a testing intermediate step. The company's long-term goal is to create AI that can design completely new experiments and deliver original scientific insights whose outcomes no one knows in advance.

Lilith's verdict

Inherent shows that universal models have hit their ceiling in narrow domains. When you send an agent to replicate science, you don't need a giant that can write sonnets and code in ten languages, but a focused expert who isn't afraid to ask a third party for a tool.