Inherent, a UK AI startup founded by several former Google DeepMind members, has formally introduced Faraday, an AI research agent built for scientific work. The company says Faraday beat OpenAI GPT-5.5 and Anthropic Claude Opus 4.8 in a targeted test focused on independently reproducing results from scientific papers, even though the system runs on a 27-billion-parameter Qwen 3.6 model.
Inherent’s pitch is straightforward: instead of only scaling up foundation models, the company wants to see whether reinforcement learning, tool use, and extended autonomous work can turn a relatively small model into a specialist agent for a defined domain.
Faraday is meant to rerun research, not just answer questions about papers
The agent is named Faraday after British scientist Michael Faraday. Inherent said that in 1821, while writing a review of work in electromagnetism, Faraday chose to reproduce earlier research himself. During that process, he found that a current-carrying wire could move around a magnet, a result that later became an important starting point for the development of the electric motor.
That process is what Inherent wants AI to mimic. The company is not aiming at a system that reads a paper and answers multiple-choice questions about it. It wants an agent that can redo the work.
To train for that, Inherent created a task suite called Replica. The initial version includes 310 tasks drawn from 100 papers in machine learning and AI for Science, covering areas such as natural language processing, materials science, and weather forecasting. The agent must reproduce research figures from those papers under limited time and compute, and it does not get access to the original figures in advance.
That forces the system to work through the same kinds of decisions a researcher would face: what exactly the authors were trying to show, which experiments need to be run, which ones deserve scarce compute, and how to revise the plan when the output does not match.
Edward Hughes says reproduction is a core part of scientific training
Edward Hughes, Inherent’s co-founder and chief scientist, told TechCrunch that reproducing papers is already a central part of scientific training, saying that “many PhD students actually start there.”
That framing matters because Inherent is positioning Faraday as a system trained to execute a research process under practical constraints, not as a paper assistant that simply produces answers.
Replica results put Faraday ahead of GPT-5.5 and Claude Opus 4.8 on average
The model size is a major part of the announcement. Faraday is built on a 27B-parameter Qwen 3.6 model, but Inherent’s Replica results say it delivered better average performance on paper reproduction tasks than Claude Opus 4.8 and GPT-5.5.
Inherent said Claude Opus 4.8 and GPT-5.5 were tested through Claude Code and Codex, respectively, with reasoning or thinking effort set to very high. According to the company, Faraday produced more faithful reproductions across different types of papers and was less likely to fail on newer research.
OpenAI has previously explored a similar direction through PaperBench, which asked an AI agent to understand a paper from scratch, write code, and run experiments. In that earlier test, OpenAI’s best agent reached an average reproduction score of just 21% and still did not beat the benchmark set by machine learning PhD students. That gap highlights how much harder it is to actually redo research than to answer questions about a paper.
Inherent says the real target is “research taste”
Inherent argues that getting an experimental result right is not enough. The company wants Faraday to learn what it calls Research Taste: which questions are worth studying, which experiments are worth running, and how to design good experiments under limited resources.
That is difficult to reduce to a fixed answer key. In response, Inherent uses long-horizon reinforcement learning, rewarding the model based on outcomes instead of prescribing each step directly. The company also built an evaluation framework and used human researchers to check how closely AI judging aligned with expert judgment.
Hughes said the company has treated an AI Scientist agent as its “North Star” from the start, and that research taste is an important piece of that goal. He also drew a line between Inherent’s work and more common products focused on helping users search or read papers.
Startup emerged from stealth after a $50 million seed round
Inherent is based in London and was co-founded by Edward Hughes, Louis Kirsch, Kaloyan Aleksiev, and Tantum Collins. The founding team includes multiple former Google DeepMind members.
A few weeks before unveiling Faraday, the company came out of stealth and closed a $50 million seed round. Inherent plans to expand its team to about 20 to 25 people by the end of the year. Beyond AI Scientist systems, the company said its research ambitions also extend to world models.