Inherent's Faraday Agent Outperforms Frontier Models in Research

Inherent, a London-based laboratory founded by former Google DeepMind researchers, claims its new AI agent has surpassed much larger frontier models in a specialized scientific task. The agent, named Faraday, demonstrated an ability to independently reproduce the findings of published scientific papers without being provided the answers in advance.

While many AI startups focus on general reasoning or creative writing, Inherent is targeting the core of the scientific method. The company recently emerged from stealth following a $50 million seed round to pursue a goal that goes beyond mere verification: the autonomous discovery of new scientific knowledge.

Replicating existing research is a foundational step for any scientist. As Inherent cofounder and chief scientist Edward Hughes noted, this process mirrors the training of PhD students who must first master established results before contributing original work. However, the technical achievement lies in the efficiency of the underlying architecture.

Efficiency and Scale

Faraday achieves these results using a significantly smaller footprint than its competitors. According to TechCrunch, the agent runs on Qwen 3.6, a model with only 27 billion parameters. This stands in stark contrast to the massive, compute-heavy architectures of Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5.

In the current landscape of artificial intelligence, the industry has largely equated performance with parameter count. Inherent is challenging this correlation by focusing on what Hughes calls research taste. This metric evaluates not just whether an agent reaches the correct conclusion, but whether it demonstrates the intuition and methodological rigor required for high-level scientific inquiry.

By optimizing for this specific type of reasoning, Faraday manages to outperform frontier-scale systems while utilizing a fraction of the computational resources. This approach suggests that for specialized domains like science and engineering, architectural intelligence may eventually supersede raw scale.

Contextualizing the Shift

The emergence of Inherent follows a broader trend of talent migrating from established giants to specialized ventures. We have seen similar movements where researchers leave large labs to focus on applying artificial intelligence to specific scientific bottlenecks. This shift highlights a growing realization that general-purpose models, while impressive, often lack the precision required for rigorous experimental loops.

As these specialized agents become more capable, the barrier to entry for complex scientific experimentation may lower. If an agent can reliably replicate a study, it can eventually be tasked with proposing the next logical experiment, effectively automating parts of the discovery pipeline. This moves the technology from a passive tool to an active collaborator in the lab.

However, the transition from replication to discovery remains an unproven leap. While Faraday has shown it can follow the breadcrumbs of existing literature, the ability to navigate the unknown requires a level of creative hypothesis generation that has yet to be standardized in any benchmark. The industry is watching to see if this efficiency can translate into genuine scientific breakthroughs.

FAQ

What is Faraday?
Faraday is an AI agent developed by Inherent that specializes in reproducing the results of scientific papers to verify research accuracy.

How does Faraday compare to GPT-5.5?
Faraday uses a 27B parameter Qwen 3.6 model, making it significantly smaller and more computationally efficient than the much larger GPT-5.5.

What is research taste in AI?
Research taste refers to an agent's ability to demonstrate scientific intuition and methodological rigor rather than just achieving high accuracy scores.