Inherent’s Faraday AI agent beats larger Anthropic and OpenAI models at paper replication using a 27B parameter model

Inherent released Faraday, an AI agent that outperforms larger models from Anthropic and OpenAI on paper replication using a smaller Qwen 3.6 27B parameter model, showing that size is not the sole determinant of performance. The agent was evaluated on its ability to independently reproduce findings from published scientific papers without being given the answer in advance. This task mirrors a standard training exercise for human scientists, indicating a meaningful benchmark for AI reasoning. By beating frontier-scale systems, Inherent highlights a shift toward efficiency in model design.

Adopting Faraday’s approach reduces computational and training expenses because a 27-billion-parameter model requires far less compute than frontier-scale counterparts. Inherent avoids building its own coding tools by leveraging OpenAI’s GPT-5.5 Codex, which lowers development overhead and mirrors how human scientists reuse existing software. The reinforcement-learning framework rewards successful outcomes instead of encoding explicit rules, cutting the need for extensive rule-based engineering. Together, these factors suggest a lower barrier to entry for teams seeking high-performing research agents.

Faraday’s current benchmark is limited to replicating known results, and the material notes that Inherent’s ultimate goal is to discover new knowledge, which the agent has not yet demonstrated. Reliance on an external coding tool (OpenAI’s GPT-5.5 Codex) may restrict the agent’s autonomy in fully self-directed research workflows. The reinforcement-learning signal for “research taste” is presented as a bet that it will generalize across scientific fields, but its transferability remains unproven. Consequently, the approach may stall when faced with novel hypothesis generation or interdisciplinary problem-solving.

Inherent’s team consists of a dozen employees with plans to expand to about 20-25, keeping the organization relatively small and potentially limiting rapid scaling. All staff work in person from a King’s Cross office, reflecting a culture that values co-location but may constrain hiring from geographically dispersed talent pools. Founder Edward Hughes has publicly criticized the U.K. garden-leave practice, highlighting a hiring barrier that could affect the startup’s ability to attract DeepMind alumni. These organizational and contextual factors shape the practical costs and feasibility of adopting Inherent’s technology at scale.