London AI Lab Inherent Says Its Agent Outperformed Models From Anthropic and OpenAI
AI

London AI Lab Inherent Says Its Agent Outperformed Models From Anthropic and OpenAI

TechNews Editorial
TechNews EditorialAug 23, 2026 · 2 min read
Share

London artificial intelligence lab Inherent stated that its new agent outperformed much larger models from Anthropic and OpenAI. The startup achieved this using a fraction of the size. Inherent recently emerged from stealth with a fifty million dollar seed round. The British firm shared details about its AI agent named Faraday.

Faraday successfully reproduced findings from published scientific papers independently without prior knowledge of the answers. Cofounder and chief scientist Edward Hughes noted that paper replication is a standard training exercise for human PhD students. Hughes emphasized that the method of building the system was more significant than simply beating rival agents.

Faraday runs on a tiny model called Qwen 3.6 with twenty-seven billion parameters. By contrast, Anthropic uses Claude Opus 4.8 and OpenAI uses GPT-5.5. Those frontier-scale systems are much larger and more expensive to train. Inherent evaluated Faraday on research taste, measuring its instinct for experiment design.

The company relies on reinforcement learning to teach this taste. Inherent wants agents to act like collaborative teammates who pursue curious experiments. The dozen employees work out of an office in King's Cross. Hughes criticized the practice of garden leave in the United Kingdom which blocks departing employees from joining rivals.

Inherent plans to grow its headcount to twenty or twenty-five people by the end of the year.

Related Stories