A $50 million seed round, a team of twelve, and a model less than a tenth the size of its competitors. According to TechCrunch, London-based Inherent claims its AI agent Faraday has outperformed Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 on a specific scientific task: independently reproducing the findings of published research papers without being shown the answers first.
The task is paper replication, which sounds narrow until you consider what it actually tests. It’s not retrieval or summarization. It requires the model to understand what a study did, reason about how to repeat it, and get the right result. Cofounder and chief scientist Edward Hughes points out that human PhD students typically start here too. It’s a baseline for scientific competence, not a parlor trick.
What makes this worth paying attention to is the hardware gap. Faraday runs on Qwen 3.6, a model with 27 billion parameters. Claude Opus 4.8 and GPT-5.5 are frontier-scale systems, almost certainly running on models orders of magnitude larger, with training costs to match. Beating them at a structured scientific reasoning task on a fraction of the compute is the kind of efficiency result that makes investors and competitors both sit up.
But Inherent’s own team says the score wasn’t the point. The method was. Rather than training Faraday on curated examples of scientific methodology, the company used reinforcement learning, a reward-based approach that lets the model figure out good strategies through trial and feedback rather than following prescribed rules. The bet is that this generalizes better across scientific fields than supervised training on domain-specific data.
Inherent also made a deliberate architectural choice that reflects a broader philosophy. Instead of building its own code execution tool, it had Faraday use OpenAI’s GPT-5.5 Codex, the same way working scientists rely on existing software. That kind of pragmatism is rare in a space where every lab tends to build everything in-house.
The longer-term goal is more ambitious than replication. Inherent wants to build AI that can generate new scientific knowledge, not just verify old results. Hughes describes the target behavior as an agent with “research taste,” meaning genuine instinct about which experiments are worth running and how to design them well. Teaching that is a harder problem than benchmarking accuracy, and it’s where reinforcement learning becomes central to the approach.
The company is based in King’s Cross, a part of London that Google DeepMind’s presence helped turn into a genuine AI cluster. Hughes has also been vocal about ending “garden leave,” the UK practice of barring departing employees from joining competitors for months after they resign. It’s a real structural disadvantage compared to US hiring norms, and one he says affected him personally before starting Inherent with three other DeepMind alumni.
Inherent currently has twelve employees and plans to reach twenty to twenty-five by year end. With DeepMind navigating internal uncertainty following changes to Demis Hassabis’s role, a well-funded London startup with credible research results and a clear scientific thesis could become an appealing next stop for researchers weighing their options.
