Inherent's 27B Model Faraday Outperforms Anthropic and OpenAI

What stands out even more is the size of the model running underneath Faraday. As Inherent explains on its own research page, Faraday is built on Qwen 3.6, a relatively small model with 27 billion parameters. Parameter count is generally used as a rough proxy for a model's size and training cost, and frontier models are typically understood to be far larger than that.

Who is Inherent

Inherent is one of the lower-profile startups founded by Google DeepMind alumni. While better-funded rivals have yet to show much in the way of concrete output, this team has quietly started putting results on the table. It's only been a few weeks since the company emerged from stealth with a $50 million seed round, and now it's already sharing these Faraday results.

Edward Hughes, co-founder and chief scientist, acknowledges that reproducing papers can look like a party trick, but he argues the task mirrors the standard training human scientists go through. "That's how most PhD students start," he said. Inherent's ultimate goal, he says, isn't just to verify existing results — it's to build AI that can discover new scientific knowledge.

How Faraday was built mattered more than winning

Hughes says beating other AI systems wasn't really the point. What mattered more was how it got there. Inherent's bar for success went beyond raw accuracy, too. It wasn't enough for Faraday to simply reproduce a result — Inherent wanted it to develop a sense of which experiments are worth running and how to design them well, something the company calls "research taste."

That kind of judgment is hard to teach through explicit rules, so Inherent leans on reinforcement learning instead — rewarding good outcomes rather than trying to hard-code the scientific process itself, on the theory that reward-based learning will generalize better across different scientific fields. "We're always drawn back to the north star of building an AI scientist agent and instilling it with taste," Hughes said.

That same principle shaped what Inherent chose not to build. Rather than developing its own coding tools, Faraday runs on OpenAI's GPT-5.5 Codex as-is. The company compares this to the way human scientists lean on existing software rather than building every tool themselves.

The King's Cross office — and the "garden leave" problem

All 12 of Inherent's employees work in person out of an office in London's King's Cross — a once-rundown district that's become one of the world's AI hubs since Google DeepMind set up shop there. "I believe London is where it should happen," Hughes said.

At the same time, Hughes has been vocal about wanting to scrap "garden leave," a common practice in the UK that bars departing employees from joining a competitor or starting their own company for months after leaving a job. Critics say this puts UK startups at a disadvantage versus the US, where researchers generally aren't bound by such restrictions, giving American startups an edge in recruiting talent. In a post on X, Hughes wrote: "This is a personal view, not a company one, but I struggled with garden leave myself."

Hughes eventually got past that restriction and co-founded Inherent with two other Google DeepMind alumni and a fourth co-founder. The company shows no sign of slowing down — it plans to grow to 20-25 people by the end of the year. With ambitions in world models as well, some see Inherent as an attractive landing spot for Google DeepMind staff unsettled by Demis Hassabis's new role there.

Model size comparison

System	Developer	Notes
Faraday (built on Qwen 3.6)	Inherent	27B parameters, specialized for paper-replication tasks
Claude Opus 4.8	Anthropic	Frontier-class, parameter count undisclosed
GPT-5.5	OpenAI	Frontier-class, parameter count undisclosed

Editor's take

The real headline here isn't that Inherent won — it's that a much smaller model won. The AI industry has largely moved toward scaling up model size and compute to push performance higher, and Inherent just produced a counterexample. A 27-billion-parameter model beating frontier-class systems suggests that how well a problem is framed and how well a model is trained can matter more than sheer size.

That reading gets reinforced when you set this alongside other recent industry signals. As NVIDIA recently demonstrated, agent performance often hinges less on the underlying model than on the harness around it — how a task gets broken down and what tools get bolted on. Inherent's decision to skip building its own coding tools and just use OpenAI's Codex fits the same logic: focus on what you do best, and borrow the rest rather than reinventing it.

There's a practical takeaway for AI teams here, too. Teams wary of large-model API costs might get better returns from narrowly defining a task and fine-tuning a small model with reinforcement learning for it, rather than defaulting to a frontier model. That said, this result is confined to the narrow task of paper replication, and Inherent itself admits that genuine scientific discovery is a much harder next step.

Over the coming months, the real test will be whether Inherent can move past this narrow task and actually design and produce novel experimental results on its own. That's when we'll find out whether "research taste" was a marketing phrase or a real capability.