Inherent, a London-based AI laboratory founded by Google DeepMind alumni, released Faraday, a 27-billion-parameter agent that successfully replicated findings from published scientific papers without prior knowledge of the answers—outperforming Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5, which operate at vastly larger scales.
The achievement hints at a potential shift in AI development economics. Parameter count typically correlates with training costs and inference expense. If smaller, specialized models can match or exceed larger generalist systems on complex reasoning tasks, it could compress the unit economics of high-performance AI deployment and pressure cloud margins for providers like OpenAI and Anthropic.
Inherent emerged from stealth weeks ago after closing a $50 million seed round. The startup built Faraday on a Qwen 3.6 base model and deliberately outsourced coding tasks to OpenAI's GPT-5.5 Codex—mirroring how human scientists rely on existing tools rather than building every component internally.
Cofounder and chief scientist Edward Hughes said the company's focus was less on winning a benchmark and more on instilling "research taste" in Faraday—an intangible ability to identify valuable experiments and design them. Paper replication, Hughes noted, is standard training for PhD students; the real test will be discovery.
Inherent trained Faraday using reinforcement learning, a reward-based approach designed to generalize across scientific domains. The method sidesteps the need for explicit rules or massive supervised datasets, potentially reducing the compute footprint required to achieve strong performance in specialized reasoning tasks.
Inherent's north star remains building AI agents that can contribute to genuine scientific discovery. The early efficiency signal sets a benchmark: in narrow, high-value domains, smaller models with focused training regimes may outcompete larger, generalist systems—a finding that could reshape how capital flows in AI infrastructure over the next two to three years.
