Inherent, founded by DeepMind alumni, says its AI agent Faraday just outperformed larger, more established models from Anthropic and OpenAI at a specific task: independently reproducing the findings of published scientific papers. The London based startup, which emerged from stealth in May 2026 with a $50 million seed round, built Faraday on top of a comparatively small open model rather than a frontier scale system of its own.
Cofounder and chief scientist Edward Hughes told TechCrunch that beating larger rivals wasn't really the point. What mattered more, he said, was how Inherent trained the agent to develop what he calls research taste: judgment about which experiments are worth running in the first place, not just the ability to check a known answer.
The release is a small but concrete data point from a startup that has kept a low profile compared with better funded peers founded by other former DeepMind researchers. It also puts a number on the table that's easy to cite: Faraday runs on a 27 billion parameter model, far smaller than the frontier systems it was measured against.
What Inherent Actually Built
Inherent is a London based AI lab founded by a group of Google DeepMind alumni, operating out of an office in King's Cross, the neighborhood DeepMind's presence helped turn into one of the world's most active AI hubs. The company emerged from stealth in May 2026 with a $50 million seed round and has kept a lower profile than some other DeepMind alumni startups, according to reporting from TechCrunch.
Its first public product is Faraday, an AI agent designed to independently reproduce the results of published scientific papers without being told the answer in advance. Inherent frames this as a training exercise rather than an end goal. Cofounder and chief scientist Edward Hughes compared it to how human scientists get started in a field. "Many PhD students actually start by doing this," he told TechCrunch. The company's longer term ambition is an AI system that can help discover new scientific knowledge, not just verify results that are already known.
How Does Faraday Compare to Claude Opus 4.8 and GPT-5.5?
Inherent says Faraday outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at the paper replication task, despite running on a much smaller base model. Faraday is built on Qwen 3.6, an open model projected to deliver approximately 27 billion parameters (pending official confirmation), a fraction of the size typically associated with frontier scale systems like Claude Opus 4.8 and GPT-5.5, whose exact parameter counts haven't been publicly disclosed by Anthropic or OpenAI. Parameter count is a rough proxy for a model's size and, often, its training cost, so a smaller model matching or beating larger ones on a specific task is the kind of result that tends to interest investors and researchers watching efficiency gains in the space.
Inherent's own bar for success went beyond raw accuracy. The company also wanted Faraday to show what Hughes calls research taste, an instinct for which experiments are worth designing and running well, rather than simply arriving at a correct number.
Why Inherent Chose Reinforcement Learning Over Rule Based Training
Teaching a model something as subjective as research taste is difficult to specify with hard coded rules, which is part of why Inherent leaned on reinforcement learning, a training method that rewards a system for good outcomes instead of listing out instructions for it to follow. Rather than training Faraday primarily on the study of how science itself is conducted, Inherent bet that a reward based approach would generalize better toward its longer term goal of agents that can contribute across many scientific fields, not just replicate results in one.
That approach also shaped what Inherent decided not to build. Instead of developing its own coding tool for Faraday, the company had the agent use OpenAI's GPT-5.5 Codex, treating it the way a human scientist might lean on existing software rather than building every tool from scratch. It's a detail that cuts against a common assumption in the AI research race, that outperforming a rival on a benchmark means avoiding that rival's technology altogether. Inherent's approach suggests a more research focused view, borrowing infrastructure where it makes sense and reserving its own training effort for the harder, less well defined problem of judgment.
The result also lands at a moment when questions about how agentic AI systems are outpacing enterprise readiness are becoming more pressing across the industry, since agent performance on narrow benchmarks doesn't always translate into reliable real world deployment.
The Team Behind Faraday
Inherent was started by Hughes alongside three other cofounders: Louis Kirsch, Kaloyan Aleksiev, and Tantum Collins. The company currently employs about a dozen people, all working in person out of its London office, and plans to grow headcount to roughly 20 to 25 by the end of the year.
Hughes has been candid that reaching this point took working around a UK specific obstacle. He has publicly called for an end to "garden leave," the common UK practice of barring departing employees from joining or starting a rival company for months after they resign, a restriction that researchers in the US generally don't face. Hughes said the position is a personal one rather than an official company stance, and that he was personally affected by the garden leave problem before starting Inherent.
That talent dynamic matters beyond one founder's story. With Demis Hassabis taking on a new role at Google DeepMind and some staff reportedly unsettled by the change, and with senior researchers already leaving Google to start their own ventures, Inherent's hiring push could make it an attractive landing spot for DeepMind staff considering a move. It's part of a broader pattern of AI agents being measured against increasingly demanding benchmarks, similar to how Nvidia's AVO harness was evaluated against the ARC-AGI-3 benchmark earlier this year, and it feeds into a larger, harder question the field hasn't settled: whether tools like Faraday move AI meaningfully closer to accelerating scientific discovery itself, a challenge our earlier coverage found remains far from solved even in high stakes fields like cancer research.
How Faraday Stacks Up Against the Models It Beat
Key Takeaways
- Inherent, a London startup founded by Google DeepMind alumni, says its AI agent Faraday outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at independently reproducing published research findings.
- Faraday runs on Qwen 3.6, a 27 billion parameter open model, far smaller than the frontier systems it was compared against.
- Inherent trained Faraday using reinforcement learning to develop research taste, judgment about which experiments are worth running, rather than optimizing for accuracy alone.
- The company emerged from stealth in May 2026 with a $50 million seed round, has about a dozen employees, and plans to grow to 20 to 25 people by the end of the year.
Frequently Asked Questions
Who were the founders of DeepMind?
DeepMind was founded in 2010 by Demis Hassabis, Shane Legg, and Mustafa Suleyman. Google acquired the company in 2014, after which it eventually became Google DeepMind.
Was Elon Musk an investor in DeepMind?
Yes. Musk made an early stage investment in DeepMind before Google's 2014 acquisition, reportedly so he could keep closer tabs on the pace of AI progress rather than for a financial return. He also reportedly tried to prevent Google from acquiring the company outright, an effort that didn't succeed.
Why did Mustafa Suleyman leave DeepMind?
Suleyman stepped away from his role leading DeepMind's applied AI division in 2019 following internal complaints about his management style, which DeepMind's leadership later acknowledged in a message to staff. He moved to a policy role at Google that same year, left Google in 2022 to become a venture capitalist, then co-founded Inflection AI before joining Microsoft to lead its AI efforts.
Did DeepMind win a Nobel Prize?
Yes. Demis Hassabis and John Jumper won the 2024 Nobel Prize in Chemistry, shared with David Baker, for work on AlphaFold, DeepMind's system for predicting protein structures.
Key Takeaways
- Inherent, a London startup founded by Google DeepMind alumni, says its AI agent Faraday outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at independently reproducing published research findings.
- Faraday runs on Qwen 3.6, a 27 billion parameter open model, far smaller than the frontier systems it was compared against.
- Inherent trained Faraday using reinforcement learning to develop research taste, judgment about which experiments are worth running, rather than optimizing for accuracy alone.
- The company emerged from stealth in May 2026 with a $50 million seed round, has about a dozen employees, and plans to grow to 20 to 25 people by the end of the year.
FAQ
Who were the founders of DeepMind?
DeepMind was founded in 2010 by Demis Hassabis, Shane Legg, and Mustafa Suleyman. Google acquired the company in 2014, after which it eventually became Google DeepMind.
Was Elon Musk an investor in DeepMind?
Yes. Musk made an early stage investment in DeepMind before Google's 2014 acquisition, reportedly so he could keep closer tabs on the pace of AI progress rather than for a financial return. He also reportedly tried to prevent Google from acquiring the company outright, an effort that didn't succeed.
Why did Mustafa Suleyman leave DeepMind?
Suleyman stepped away from his role leading DeepMind's applied AI division in 2019 following internal complaints about his management style, which DeepMind's leadership later acknowledged in a message to staff. He moved to a policy role at Google that same year, left Google in 2022 to become a venture capitalist, then co-founded Inflection AI before joining Microsoft to lead its AI efforts.
Did DeepMind win a Nobel Prize?
Yes. Demis Hassabis and John Jumper won the 2024 Nobel Prize in Chemistry, shared with David Baker, for work on AlphaFold, DeepMind's system for predicting protein structures.
