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

What happened

British AI lab Inherent, founded by former DeepMind researchers, has launched Faraday, an AI agent designed to replicate published scientific research. The company says Faraday outperformed comparable systems from Anthropic and OpenAI on the task of reproducing findings from existing papers, positioning the tool as an early step towards AI-assisted scientific discovery.

Faraday is pitched as an AI "teammate" rather than a simple chatbot: it is built to work through the mechanics of experimental replication — reading a paper, understanding its methodology, and attempting to reproduce its results — a task that requires sustained reasoning rather than a single-shot answer.

Why it matters

Replication is one of the more laborious, unglamorous parts of scientific work, and it is also a useful proving ground for AI agents because success or failure is relatively easy to verify against a known result. A system that can reliably replicate published research suggests a path towards AI agents that can eventually assist with hypothesis generation, experimental design, or first-pass validation — work currently bottlenecked by scarce researcher time.

For organisations investing in AI, the story is less about Faraday itself than about what it signals: specialised agents built around a single, verifiable task can outperform general-purpose frontier models on that task, even when built by a much smaller lab. That has implications for how enterprises think about buying or building AI capability — the biggest model is not always the best tool for a narrow, high-value job.

The Renascence take

The headline comparison to Anthropic and OpenAI will get attention, but the more interesting signal is about specialisation versus scale.

What Inherent is really demonstrating is that a small, focused team can beat general-purpose frontier labs at a narrowly defined, verifiable task — and that lesson applies well beyond science. Most organisations chase the biggest, most capable model available, when the better move is often matching a tightly scoped agent to a tightly scoped job. Leaders evaluating AI for research, quality assurance or knowledge work should ask less "which model is smartest" and more "which task is narrow and verifiable enough that a purpose-built agent will outperform a generalist one." That reframing changes procurement, vendor selection and where AI investment actually pays off.