Inherent's small Faraday agent claims research replication win over frontier AI models

The claim

Inherent, a London AI lab founded by former Google DeepMind staff, says its newly released agent Faraday has outperformed much larger models from Anthropic and OpenAI at independently reproducing the results of published scientific papers, without being given the answers in advance. The claim, reported by TechCrunch, comes from the company itself and has not been independently verified.

The comparison set Faraday against Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5, both frontier-scale systems. According to TechCrunch, Faraday runs on Qwen 3.6, a model with 27 billion parameters — a fraction of the size, with parameter count serving as a rough proxy for both scale and training cost.

A startup fresh out of stealth

Inherent emerged from stealth only weeks ago with a $50 million seed round, and is led by cofounders Louis Kirsch, Kaloyan Aleksiev, Tantum Collins and Edward Hughes. Hughes, the chief scientist, frames replication as a training exercise familiar to human researchers, noting that many PhD students begin their careers by reproducing existing work before producing anything original. The lab's longer-term ambition is AI that discovers new scientific knowledge rather than merely checking old results.

Teaching an agent research judgement

Accuracy was not the only bar. Hughes told TechCrunch that Inherent also wanted Faraday to demonstrate what it calls "research taste" — an instinct for which experiments are worth running and how to design them well. Hughes said the more interesting part of the exercise was not beating frontier agents, "which of course we liked", but how the team built the system.

That method is reinforcement learning, which rewards the agent for good outcomes instead of prescribing explicit rules. Rather than grounding training primarily in the formal study of how science itself is conducted, Inherent leans on this reward-based approach, betting it will generalize better toward agents capable of contributing across many scientific fields.

The company has also drawn boundaries around what it will not build. Inherent developed no coding tool of its own and had Faraday use OpenAI's GPT-5.5 Codex, mirroring the way human scientists adopt existing software rather than writing everything from scratch.

Hughes said the template for the agent is a particular kind of colleague: one who comes back saying they got curious, ran some experiments, and asks what you think of the results. The aim, he said, is to avoid agents that simply tell users what they want to hear.

London hiring push

Inherent's roughly one dozen employees all work in person out of an office in King's Cross, the London district that DeepMind's presence helped transform into a major AI hub. The company plans to grow to about 20 to 25 people by the end of the year.

Hughes is bullish on London's density of AI talent, but he has separately criticised the UK's "garden leave" practice, which bars departing employees from joining or starting a rival company for a period after resigning — a restriction he says American researchers generally do not face. He described this as a personal view rather than a company position, and said he was personally affected by it. TechCrunch also notes that with Demis Hassabis taking on a new role and some DeepMind staff unsettled, Inherent's hiring push could make it an appealing destination for DeepMind employees weighing a move.

Why it matters

If a 27-billion-parameter model, trained with the right reward scheme, can beat frontier-scale systems at autonomous paper replication, it suggests that targeted reinforcement learning may matter as much as raw scale for agentic scientific work. The claim is also a marker on the path from verification to discovery: replication is the warm-up, and Inherent's stated destination is AI that generates new scientific knowledge. The caveats are significant — the benchmark is narrow, the evaluation was run by the startup itself, and neither Anthropic nor OpenAI has responded publicly — so the result is best read as a directional signal about where small, deliberately trained agents are heading rather than a settled verdict on model rankings.