DeepMind Spinout Inherent Claims Its Compact AI Agent Beats Claude and GPT-5.5 at Research Replication

London-based Inherent, a startup founded by former Google DeepMind researchers, claims its new AI agent has surpassed larger competitors from Anthropic and OpenAI on a challenging scientific task—all while running on a dramatically smaller model.

The company emerged from stealth just weeks ago with a $50 million seed round and has now released Faraday, an AI agent designed to reproduce published research findings independently. According to Inherent, Faraday managed to outperform Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at this task, despite running on the Qwen 3.6 model—which contains only 27 billion parameters, roughly 100 times smaller than the frontier systems it beat.

Why Paper Replication Matters#

Reproducing existing research may seem like a modest accomplishment, but it mirrors a standard early step in the scientific process itself. Edward Hughes, Inherent's cofounder and chief scientist, noted that many PhD students begin their careers by replicating prior work. Rather than viewing this as a party trick, Hughes emphasized that the method behind the achievement was equally significant—the lab's approach to building the agent matters as much as the results it produces.

Building "Research Taste" Through Reinforcement Learning#

Inherent's ambition extends beyond simple accuracy. The team set out to instill what Hughes calls "research taste" in Faraday—an intuitive sense for which experiments merit pursuit and how to design them rigorously. Achieving this required moving beyond conventional training approaches. Instead of primarily teaching agents the mechanics of scientific methodology, Inherent employed reinforcement learning, a training technique that rewards systems for good outcomes rather than prescribing rigid rules.

This reward-based strategy proved flexible enough that the team avoided building proprietary tools where established alternatives work well. For coding tasks, Faraday leverages OpenAI's GPT-5.5 Codex, mirroring how human scientists rely on existing software rather than reinventing it internally.

Aiming for Collaborative AI, Not Yes-Men#

Hughes framed the vision for Faraday around genuine collaboration. Rather than producing findings that simply validate user assumptions, the goal is an AI teammate that explores independently and returns with results worth discussing—one that tells you, "I got curious about this, and I went off and did these experiments. What do you think of these results?"