A London-based AI startup founded by Google DeepMind alumni says its compact model has surpassed much larger rivals at a key scientific benchmark. Inherent’s new agent, Faraday, replicated findings from published papers without prior guidance, outperforming Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 despite running on just 27 billion parameters—far fewer than the frontier-scale systems it beat.
The company’s focus isn’t just on accuracy but on teaching the AI “research taste”—an instinct for which experiments to prioritize. Inherent used reinforcement learning to reward good outcomes rather than prescribing rigid rules, a method it believes will scale better for its long-term goal: building AI capable of original scientific discovery. Faraday also relied on OpenAI’s Codex for coding tasks, mirroring how human researchers leverage existing tools instead of reinventing them.
Inherent, which emerged from stealth in September with a $50 million seed round, plans to expand its 12-person team to 20–25 by year’s end. Cofounder Edward Hughes highlighted London’s AI talent density and criticized U.K. “garden leave” rules that delay departing researchers from joining rivals, a practice he says puts American startups at an advantage.
