Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating research

DeepMind Alumni Lab Claims Small Agent Outperforms Frontier Models

Inherent, a London-based AI lab founded by Google DeepMind alumni, announced its new AI agent, Faraday, has outperformed much larger models from Anthropic and OpenAI on the specific task of independently replicating published scientific papers. This achievement is notable because it challenges the prevailing industry assumption that superior performance requires ever-larger models. By demonstrating high-level scientific reasoning with a comparatively small model, Inherent signals a potential shift towards more efficient, specialized AI systems for complex, domain-specific tasks.

Technical Approach and 'Research Taste'

Just weeks after a $50 million seed round, Inherent detailed how Faraday operates on a 27-billion-parameter model called Qwen 3.6, a fraction of the size of frontier models like Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5. The team's co-founder, Edward Hughes, explained that their success stems from a unique training methodology. Instead of just optimizing for accuracy, they used reinforcement learning to reward the agent for demonstrating “research taste”—an instinct for designing valuable experiments. To complete its work, Faraday also uses external tools like OpenAI's GPT-5.5 Codex for coding, mirroring the workflow of a human scientist who leverages existing software rather than building everything from scratch.

• Agent Name: Faraday
• Company: Inherent (London, UK)
• Base Model: Qwen 3.6 (27 billion parameters)
• Core Task: Autonomous replication of published scientific papers
• Key Differentiator: Trained with reinforcement learning to develop "research taste"
• Outperformed: Anthropic Claude Opus 4.8 and OpenAI GPT-5.5 on the specified task

Ecosystem Impact and Talent Hub

Inherent's approach could influence a segment of the market focused on building capable, cost-effective agents rather than monolithic, general-purpose models. The company's progress reinforces London's status as a key AI hub, a position cultivated in part by the presence of Google DeepMind. With plans to nearly double its team by the end of the year, Inherent could become a prime destination for AI talent looking to leave larger labs. Co-founder Edward Hughes has also publicly criticized the UK's “garden leave” employment practice, which he argues slows down innovation by preventing researchers from quickly joining new ventures.

Inherent’s work with Faraday suggests the next significant advance in AI may not come from the largest model, but from smaller, specialized agents trained with methods like reinforcement learning to develop nuanced, domain-specific judgment.