In an industry where every month somebody announces a bigger model than yesterday's, the London lab Inherent says it did the opposite - an agent running on a model with 27 billion parameters that outperforms systems incomparably larger than it.

The company was founded by former Google DeepMind people and until recently worked quietly. A few weeks ago it stepped out of the shadows with a 50 million dollar seed round, and now it is showing what it was building: an agent called Faraday, whose task is to independently reproduce the results of published scientific papers - without being told in advance what the correct answer is.

According to the company, on that task Faraday outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 - both so-called frontier-class models, significantly larger. Parameters are not a perfect measure, but they roughly indicate what it costs to train a model and what it costs to run it. In this case the difference is not cosmetic.

Not the result, but the method

"What we found most interesting was not so much the result that we beat those frontier agents - which we certainly liked - but the way we got there," says Edward Hughes, co-founder and chief scientist at Inherent.

Replicating other people's work sounds like a modest ambition for a company that says it wants to build artificial intelligence capable of discovering new scientific knowledge. Hughes defends it with an analogy from academia itself: "A lot of PhD students actually start out doing exactly that."

The bar the company set itself was not only accuracy. They wanted the agent to show what they call "research taste" - an instinct for which experiment is even worth doing and how to design it. That is hard to teach through rules, so Inherent relies on reinforcement learning: the model is rewarded for a good outcome rather than being prescribed a procedure.

There is also a decision that says more than all the rhetoric - Inherent did not build its own coding tool. Instead it let Faraday use OpenAI's GPT-5.5 Codex. It builds on a competitor, just as a scientist uses somebody else's software rather than writing it from scratch.

London, twelve people and one British obstacle

The whole team numbers twelve employees and works out of an office in King's Cross - once a run-down part of London that Google DeepMind's presence turned into one of the world's centres for artificial intelligence. "We believe London is the place to be," says Hughes.

That same Hughes, however, openly calls for the abolition of the British practice known as gardening leave - the ban on an employee who has resigned joining a competitor or founding their own company for months. American researchers as a rule have no such obligation, which gives startups on the other side of the ocean an advantage in poaching staff. "This is a personal view, not the company's, but I was hit by the gardening leave problem," he says.

That is a detail worth noting for anyone wondering why European technology capital regularly ends up at an American address. It is not always a question of money - sometimes it is a question of six months of enforced pause paid for with a lost race.

Inherent plans to grow to twenty or twenty-five people by the end of the year. With uncertainty around Demis Hassabis's new role at DeepMind, a lab a few streets away suddenly looks like a comfortable place to land.