A London-based AI lab claims its newly released agent has outperformed much larger models from Anthropic and OpenAI at a key scientific task while running on a fraction of their computational power. Inherent, a startup founded by former Google DeepMind researchers, announced this week that Faraday, its AI agent, surpassed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at independently reproducing the findings of published scientific papers without advance knowledge of the results. The British company emerged from stealth just weeks ago with $50 million in seed funding.
Faraday operates on a model called Qwen 3.6 with just 27 billion parameters, a stark contrast to the frontier-scale systems it was measured against. Parameters serve as a rough proxy for a model's size and training costs, making Inherent's achievement notable for its efficiency. The task itself—paper replication—involves an AI system independently verifying published scientific findings without being told the answer beforehand. Inherent set a bar beyond simple accuracy, requiring Faraday to demonstrate what it calls "research taste," meaning an instinct for which experiments are worth conducting and how to design them effectively.
Cofounder and chief scientist Edward Hughes told TechCrunch that beating competing AI systems wasn't the primary objective. "What was most interesting to us about this was not so much the result of beating those frontier agents—which of course we liked—but was actually the way we went about building this," he said. Hughes added that paper replication serves as a standard training exercise for human scientists as well, noting that "many PhD students actually start by doing this." Rather than developing its own coding tool, Inherent had Faraday use OpenAI's GPT-5.5 Codex instead, mirroring how human scientists rely on existing software rather than building everything from scratch, according to the company.
Inherent's approach centers on reinforcement learning, a training method that rewards an AI system for good outcomes rather than prescribing rules for it to follow. Teaching something as subjective as taste proves difficult, which is where this reward-based approach comes in. Rather than training its agents primarily on the study of how science itself is conducted, Inherent leans on this method because it expects the approach will generalize better to its long-term ambition: building agents capable of contributing across multiple scientific disciplines. That north star of creating an AI scientist agent guides what the company chooses not to build as well. Hughes said the goal is modeled on the kind of teammate who returns with curiosity-driven initiative, saying "I got curious about this, and I went off and I did these experiments. What do you think of these results?" Paper replication may sound like a mere party trick given Inherent's much loftier goal of building AI that can discover new scientific knowledge, not just verify old results, but it represents a crucial stepping stone toward that vision.
The startup's dozen employees all work in person from an office in King's Cross, the once-declining London neighborhood that Google DeepMind's presence helped transform into a major global AI hub. Hughes expressed confidence in London's density of AI talent, saying "we believe that London is the place to be." However, he also voiced support for ending "garden leave," the UK practice of preventing departing employees from joining or starting rival companies for months after resignation—a restriction American researchers generally don't face, giving US startups a head start on hiring talent who've left prior roles. Inherent plans to grow its headcount to approximately 20 to 25 by year's end. With Demis Hassabis's new role reportedly leaving some DeepMind staff unsettled, Inherent's hiring push could position it as an attractive destination for DeepMind employees considering a move. The efficiency gains demonstrated by Faraday suggest that smaller, more focused teams might compete with well-funded giants through architectural cleverness rather than raw computational scale. Whether reinforcement learning can truly instill scientific intuition at the level needed for genuine discovery remains the defining question for Inherent and the broader field.