Inherent's Faraday AI beats larger rivals in scientific paper replication
London startup Inherent says its Faraday agent outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at reproducing published research results.
Inherent, a London-based AI lab started by ex-DeepMind staff, announced that its Faraday agent achieved higher performance than Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 in the task of independently reproducing results from published scientific papers. Faraday runs on the comparatively small Qwen 3.6 model with 27 billion parameters, yet it also exhibited "research taste" by selecting worthwhile experiments, a capability honed through reinforcement-learning rewards.
Rather than creating its own coding engine, Faraday leveraged OpenAI's GPT-5.5 Codex, mirroring how human scientists use existing software. Co-founder and chief scientist Edward Hughes highlighted the importance of the development approach over the headline win. The firm, operating from a King’s Cross office, intends to grow its headcount to roughly 20-25 employees before the year closes, positioning itself as a potential destination for talent from DeepMind.
Why it matters
The breakthrough shows smaller AI models can match larger rivals in scientific tasks, potentially lowering costs for research automation.
In this story
