Inherent, a London AI lab founded by Google DeepMind alumni, said its AI agent Faraday outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at reproducing findings from published research papers. The company emerged from stealth in May 2026 with a $50 million seed round, according to TechCrunch.
Faraday runs on a 27-billion-parameter model called Qwen 3.6, smaller than either frontier system it was tested against. Inherent built the comparison on Replica, a benchmark of 310 tasks drawn from 100 machine learning and AI-for-science papers, each asking an agent to replicate a published figure under a limited time and compute budget, according to the company's research page.
Inherent said Faraday produced more faithful replications than the baseline models across every paper category in the test set, with particular gains in meta-learning, structural biology and materials science, and that it generalized to papers published after its training cutoff.
Cofounder and chief scientist Edward Hughes told TechCrunch the result mattered less than how Faraday was built: the agent uses reinforcement learning to develop what the company calls research taste, and calls on OpenAI's GPT-5.5 Codex as a tool rather than relying on custom coding infrastructure. "Many PhD students actually start by doing this," he said, referring to paper replication as standard scientist training.
It's an unverified claim from the company that built the benchmark, worth reading as exactly that. But the underlying pattern, a small model directing a larger one as a tool instead of being replaced by it, is a cheaper way to build a capable agent if it holds up outside a test Inherent designed itself.