Inherent, a London-based artificial intelligence lab founded by former Google DeepMind employees, said that its new agent Faraday outperformed larger models from Anthropic and OpenAI on a task involving the reproduction of results from published scientific papers, without being given the answer in advance.
Faraday emerged just a few weeks after Inherent came out of stealth, alongside the company’s announcement of a $50 million seed funding round. The company says the agent was compared with Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5, two systems described in the article as larger and belonging to the advanced-model category.
A Test That Goes Beyond Matching Results
According to co-founder and chief scientist Edward Hughes, Inherent did not limit itself to measuring Faraday’s ability to reproduce results. It also wanted to test what it calls “research taste”—the ability to judge which experiments are worth carrying out and to design them appropriately. The company considers reproducing the results of published research to be common practical training for scientists, as many doctoral students begin with this task.
Faraday runs on a Qwen 3.6 model with only 27 billion parameters, according to the article, while the comparison was based on much larger models. Parameter counts are commonly used as a rough indicator of a model’s size and training costs, but the number alone is not enough to establish that one system outperforms another, particularly without complete details about the test, its data, or how the results were calculated.
Reinforcement Learning Instead of Dictating Research Steps
Inherent relies heavily on reinforcement learning, in which the system receives rewards for good results instead of being provided with a detailed set of rules for conducting scientific research. The company is betting that this approach may help its agents generalize across multiple scientific fields rather than confining them to imitating what they have learned about previous research methods.
The company says its long-term goal is to build an agent that works as an artificial intelligence scientist and contributes to the discovery of new knowledge, rather than merely verifying known results. For this reason, it did not develop its own programming tool; instead, it made Faraday use GPT-5.5 Codex, simulating researchers’ reliance on existing software rather than building every tool from scratch.
Why Does This News Matter?
If the test results are confirmed using a reproducible methodology, they could indicate that specialized agents can achieve strong performance using smaller models when supported by appropriate training and capabilities for selecting and evaluating experiments. This shifts the focus away from model size alone toward agent design, its reward mechanism, and the working environment in which it operates.
However, the source conveys the company’s claim and does not provide enough detail to independently assess the comparison. It also does not clarify the range of scientific papers or the fields included in the test. The significance of “outperformance” therefore remains limited to the specific test and does not yet establish Faraday’s ability to discover new knowledge or work effectively across all disciplines.
Inherent has about 12 employees working in person at its office in London’s King’s Cross area and plans to increase the number to about 20 or 25 employees by the end of the year. Hughes also spoke in his personal capacity about the impact of the United Kingdom’s “garden leave” restrictions on employees moving to competing companies, which could affect the lab’s ability to recruit researchers from DeepMind and elsewhere.