Inherent’s tiny agent outperforms Anthropic and OpenAI on scientific paper reconstruction

Inherent, a London-based AI lab founded by alumni of Google DeepMind, emerged from stealth a few weeks ago after raising a side round of $50 million. The announcement today says its new agent, Faraday, was able to reproduce findings from published scientific articles without being given the answers in advance, and did so more effectively than the larger models from Anthropic and OpenAI. Faraday runs on Qwen 3.6, a model with 27 billion parameters, whereas the competing systems are frontier-scale models—Claude Opus 4.8 and GPT-5.5.

Edward Hughes, co-founder, partner and chief scientist, said the benchmark win was not the primary goal. “What mattered to us was the way Inherent set a higher bar for accuracy and demanded the agent demonstrate ‘research taste’—an instinct for which experiments are worth running and how to design them properly,” he explained. He added that this quality is hard to encode with explicit rules, so the company relies on reinforcement learning, a training method that rewards good outcomes instead of prescribing fixed rules. The bet, Hughes said, is that this approach will better serve long-term objectives: agents that contribute across scientific domains.

The focus on a “north star” for a scientific agent also dictated what Inherent does not build. Rather than creating its own code-generation tool, Faraday uses the Codex of GPT-5.5, mirroring how human researchers rely on existing software instead of writing everything from scratch. Hughes described the ideal as an agent that behaves like a good teammate: “It comes back and says, ‘I got curious about this, I ran these experiments, what do you think of the results?’ rather than just telling the user what they want to hear.”

Inherent currently employs about a dozen staff, all working together in a King's Cross office—the London district that, helped in part by DeepMind’s presence, has become one of the world’s leading AI hubs. Hughes expressed optimism about London’s talent density, but also joined calls to abolish the “garden leave” practice that bars departing employees from joining or founding competitors for several months. American researchers do not face such a restriction, which gives U.S. startups an edge in attracting talent who have left previous roles.