Inherent’s Faraday AI Beats Claude and GPT-5.5 at Research Replication

Inherent, a London AI lab co-founded by four Google DeepMind alumni, says its agent Faraday has outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at independently replicating the findings of published scientific papers. The twist: Faraday runs on Qwen 3.6, a 27-billion-parameter model that is a fraction of the size of those frontier systems. The company emerged from stealth just weeks ago with a $50 million seed round and currently employs about 12 people out of an office in London's King's Cross.

What happened

Detail	Fact
Company	Inherent, London
Founders	Four Google DeepMind alumni, including cofounder and chief scientist Edward Hughes
Seed round	$50 million
Agent name	Faraday
Model powering Faraday	Qwen 3.6, 27 billion parameters
Competitors it beat	Anthropic Claude Opus 4.8, OpenAI GPT-5.5
Task	Independently replicating results from published scientific papers
Current headcount	~12, targeting 20-25 by end of year

Faraday’s benchmark task, replicating the results of scientific papers without being told the answers in advance, may sound narrow. But Edward Hughes argues it is a real scientific skill. “Many PhD students actually start by doing this,” he told TechCrunch. The point, he added, was less about beating rivals and more about the method Inherent used to get there.

That method is reinforcement learning, a training approach that rewards an AI system for producing good outcomes rather than following predefined rules. Most AI training for scientific tasks leans heavily on studying how science is conducted. Inherent is betting a reward-based approach will generalise better across many scientific fields over time.

How Faraday is different from a bigger model

Parameters are a rough proxy for model size and training cost. At 27 billion, Qwen 3.6 is far smaller than frontier models like Claude Opus 4.8 or GPT-5.5. Running a smaller model typically means lower inference costs and faster iteration, which matters if you are trying to run many experiments autonomously.

Inherent also drew a deliberate boundary around what Faraday should build itself. Rather than writing its own coding tools, Faraday uses OpenAI’s GPT-5.5 Codex for coding tasks, the same way a working scientist uses existing software rather than programming everything from scratch.

The team also set a higher bar than raw accuracy. They wanted Faraday to show “research taste”, an instinct for which experiments are worth running and how to design them well. Teaching that kind of judgment is hard with a rules-based system, which is another reason Inherent reached for reinforcement learning.

Why it matters

This result is a data point in a larger argument: that smaller, purpose-trained agents can compete with or beat general-purpose frontier models on specific professional tasks. That has real cost implications. A 27-billion-parameter model costs far less to run than a frontier-scale system, and if task-specific training closes the quality gap, the economics of deploying AI agents in research, and in other professional domains, shift considerably.

Inherent’s stated north star is an AI system that can make genuinely new scientific discoveries, not just verify old ones. Paper replication is an early milestone. But the reinforcement learning approach is the more interesting strategic bet: rather than telling the agent what good science looks like, Inherent is trying to let it learn that through experience. If it works at scale, it would be a meaningful step beyond current AI tools in the research space.

The company’s London location is also a deliberate signal. Hughes is vocal about the city’s AI talent density. He is equally vocal about “garden leave,” the UK practice of barring departing employees from joining competitors for months after they resign. Hughes told TechCrunch he was personally affected by it. He has called publicly for the practice to end, noting it gives US startups a hiring advantage that UK labs have to work around.

Our take

Self-reported benchmarks from a freshly funded startup deserve a skeptical read. Inherent has not published independent peer-reviewed results, and the comparison set (Claude Opus 4.8 and GPT-5.5 on paper replication) is one it chose itself. That said, the specific claim is testable and the methodology, reinforcement learning aimed at instilling scientific judgment rather than just accuracy, is a credible research direction. We have seen similar efficiency wins in other domains where small fine-tuned models beat larger general-purpose ones at narrow tasks. The more interesting question is whether the approach scales beyond replication into genuine discovery. That will take longer to validate.

For businesses thinking about AI integration in research-heavy workflows, the takeaway here is that model size is increasingly a poor proxy for capability on specific tasks. The right question is fit for purpose, not biggest model available. Stories like this one are worth tracking alongside broader AI news coverage as the agent landscape matures.

What to do about it

1. Watch for an independent evaluation of Faraday’s replication benchmark before drawing conclusions from Inherent’s self-reported numbers.
2. If you are deploying AI agents on professional tasks, audit whether a smaller, task-specific model might outperform a larger general one at lower cost.
3. Track how Inherent’s reinforcement learning approach performs as it moves from replication toward novel research tasks, that transition is the real test.

Source: TechCrunch · AI