Inherent AI Agent Outperforms OpenAI, Anthropic at Research Replication
London-based AI laboratory Inherent, founded by alumni of Google DeepMind, has announced that its newly launched scientific research agent, Faraday, successfully outperformed larger frontier models from Anthropic and OpenAI in independently replicating published scientific studies. Emerging from stealth with a fifty million dollar seed round, the startup demonstrates that highly capable AI researchers can operate efficiently on significantly smaller architectures. Faraday achieves these results by running on the twenty seven billion parameter Qwen 3.6 model, contrasting sharply with the substantially larger Claude Opus 4.8 and GPT 5.5 systems used for benchmark comparison. The development focuses on more than simple accuracy. Inherent engineers prioritized instilling research taste, an AI capability for evaluating experimental merit and designing rigorous protocols. The team achieved this through reinforcement learning, which rewards optimal decision-making pathways rather than relying solely on rule-based instruction or extensive historical analysis of scientific literature. This reward-based framework is designed to generalize across multiple scientific disciplines, aligning with Inherent’s broader objective of creating autonomous AI scientists capable of generating novel discoveries rather than merely validating existing work. To maintain efficiency, Faraday integrates existing infrastructure, such as OpenAI’s coding environment, rather than developing proprietary tooling from scratch. Inherent operates from a physical office in King’s Cross, London, leveraging the region’s dense concentration of artificial intelligence talent. The company currently employs approximately a dozen researchers and plans to expand its workforce to between twenty and twenty five individuals by year-end. Co-founder and chief scientist Edward Hughes emphasized the firm’s collaborative philosophy, designing Faraday to function as an independent research partner that proactively explores hypotheses and presents findings for peer evaluation. This approach mirrors modern academic methodologies, where initial training often involves replicating established studies before pursuing original inquiry. The announcement also highlights broader industry dynamics surrounding talent mobility and regional development. Hughes publicly advocated for reforming the United Kingdom’s garden leave practice, which restricts departing employees from joining competitors for extended periods. He noted that such contractual limitations disadvantage British startups relative to their American counterparts in securing specialized expertise. Additionally, recent organizational shifts at Google DeepMind, including leadership changes under Demis Hassabis, may influence researcher mobility. Inherent’s focused hiring strategy and emphasis on foundational AI scientist development position it as a potential destination for specialists evaluating opportunities within the European AI ecosystem. The startup’s ability to achieve frontier-level research replication on a fraction of the computational overhead suggests a viable pathway toward scalable, cost-effective autonomous scientific discovery.
