A tiny AI agent from a London startup just beat Anthropic and OpenAI at reading scientific papers
Inherent, founded by four Google DeepMind veterans, says its Faraday agent outperformed much larger models from Anthropic and OpenAI on a key science benchmark, despite running on a model roughly a tenth of their size.
Key points
- Inherent, a London AI startup founded by four Google DeepMind alumni, raised a $50 million seed round before releasing its first public results.
- Its AI agent, Faraday, beat Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at independently replicating published scientific research without being given the answers first.
- Faraday runs on Qwen 3.6, a 27-billion-parameter model, making it far smaller than the frontier models it outscored.
- Inherent uses reinforcement learning, a training method that rewards good outcomes rather than spelling out fixed rules, to teach Faraday what researchers call "research taste."
- The company plans to grow from 12 to roughly 20 to 25 employees by the end of 2025.
A small London lab is making a loud claim. Inherent, whose four founders all left Google DeepMind to start it, says its AI agent has beaten some of the most capable AI systems on the market at a standard scientific task, using a fraction of the computing muscle. The company came out of stealth just weeks ago with a $50 million seed round. Now it's sharing the first real look at what it has built.
What did Faraday actually do?
Faraday was set the task of replicating published scientific papers on its own, without being told the correct answers. That means reading the paper, deciding what experiments to run, and arriving at the same conclusions the original researchers did.
It beat both Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at this. First reported by TechCrunch, the result draws attention because of how Faraday pulled it off. Faraday runs on Qwen 3.6, which has 27 billion parameters, the internal numeric settings that shape how a model reasons. More parameters generally mean a bigger, more expensive model. Claude Opus 4.8 and GPT-5.5 are widely understood to operate at far larger scale.
For the company, the benchmark result wasn't even the point. "What was most interesting to us about this was not so much the result of beating those frontier agents," Edward Hughes, Inherent's cofounder and chief scientist, told TechCrunch, "but was actually the way we went about building this."
Why does paper replication matter?
It's a standard starting exercise for new scientists. "Many PhD students actually start by doing this," Hughes said. The aim isn't just accuracy; it's demonstrating you understand the process well enough to rebuild it from scratch.
Inherent's bar was higher still. Beyond accuracy, it wanted Faraday to show "research taste": an instinct for which experiments are worth running and how to design them well. Teaching something that intangible required a specific approach. Reinforcement learning rewards an AI for producing good outcomes rather than handing it a rule book, and the bet is that this generalises across many scientific fields, not just those Faraday trained on.
Faraday also uses OpenAI's GPT-5.5 Codex for coding tasks, the same way a working scientist reaches for existing software rather than writing everything from scratch.
Should you care about an agent with "research taste"?
Inherent's longer goal is an AI that discovers new things rather than just verifying old ones. Replication is the foundation, not the destination. The practical question is whether reward-based training at 27 billion parameters can generalise far enough to get there, and today's benchmark offers only one data point.
Hughes describes his ideal agent as a teammate who "got curious about this, and I went off and I did these experiments" and then asks what you think of the results. That collaborative quality is what Inherent says it's building toward. We first covered Inherent's founding on 22 August 2026, when the team was still unnamed publicly.
All 12 staff work in person from King's Cross in London. Hughes has publicly called for the UK to scrap "garden leave," the British employment clause that can bar departing workers from joining a rival for months after they quit, saying it slows UK AI startups relative to US rivals. He told TechCrunch it's a personal view, not a company position, though he was personally affected by it. The company plans to reach about 20 to 25 people before the year ends.
