--- title: "Former Google DeepMind employees unveiled Faraday, an AI agent that outperformed OpenAI and Anthropic models in scientific tests" description: "British startup Inherent unveiled Faraday, a 27-billion-parameter AI agent that outperformed OpenAI and Anthropic models in internal tests on replicating scientific research." date: 2026-08-23T17:24:46.000Z lang: en url: https://xab.info/en/posts/inherent-faraday-ai-agent-outperforms-openai-anthropic tags: [ai-agent, inherent, google-deepmind, scientific-research, reinforcement-learning] publisher: "XAB.info" --- # Former Google DeepMind employees unveiled Faraday, an AI agent that outperformed OpenAI and Anthropic models in scientific tests ![Holographic AI brain with a single eye projected from a laptop — visualization of the Faraday AI agent by Google DeepMind alumni](https://xab.info/media/2026/08/24/inherent-faraday-ii-agent-nauka/inherent-faraday-ii-agent-nauka-1.webp) ## 🎯 Key Points - Inherent (founded by former Google DeepMind staff) unveiled Faraday, an AI agent that reproduces the findings of scientific publications without prompts. - The system runs on the Qwen 3.6 model with 27 billion parameters and outperformed Claude Opus 4.8 and GPT-5.5 in internal tests. - The startup raised $50 million at an early stage and trains the agent using reinforcement learning. - Faraday uses OpenAI's GPT-5.5 Codex for programming, which makes a direct comparison with GPT-5.5 methodologically debatable. British startup Inherent, founded by former Google DeepMind staff, emerged from stealth mode just a few weeks ago and immediately revealed its key product — Faraday, an AI agent capable of independently verifying the validity of scientific publications. According to TechCrunch, the company raised $50 million in an early funding round almost simultaneously with its public announcement. The developers claim that Faraday can reproduce the findings of already published scientific studies without any prior prompts, and that in internal tests it outperformed significantly larger frontier models from competitors. ### How Faraday works and what it can do The system's key feature is that it runs on a relatively small model, Qwen 3.6, with 27 billion parameters. In its internal tests, Inherent compared Faraday against Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 — noticeably larger models. What was evaluated was not only the ability to reproduce the results of the original work, but also so-called "research intuition": the ability to independently decide which experiments are worth conducting and how to organize them. Inherent co-founder and chief scientist Edward Hughes noted that this kind of work is standard preparation for humans as well: many PhD students begin their scientific careers precisely by replicating studies that have already been carried out. ### Training method and the project's ultimate goal Faraday was trained using reinforcement learning (RL): the system receives a reward for a successful outcome rather than following pre-defined rigid rules. Hughes emphasized that for the team, this approach mattered more than beating competitors. Inherent's ultimate goal is to create an AI scientist capable not just of checking existing conclusions, but of acquiring new scientific knowledge. At the same time, the startup is not trying to build every tool in-house: for programming, for example, Faraday uses OpenAI's GPT-5.5 Codex, which Inherent compares to ordinary researchers using off-the-shelf software. ### Contradictory data It is important to note that the claimed superiority of Faraday over Claude Opus 4.8 and GPT-5.5 is based on the startup's own internal tests, not on an independent public benchmark, so the figures and conclusions cannot yet be considered verified by a third party. Moreover, because the agent relies on OpenAI's GPT-5.5 Codex for programming, a direct comparison of its results with a "pure" GPT-5.5 is methodologically debatable: part of the computational work is in fact performed by a tool from the same vendor being compared against. The developers, in turn, stress that this is an acceptable practice, analogous to using ready-made software in real science. ### Team, location, and stance on garden leave Inherent currently has around 12 employees in London's King's Cross district, with plans to expand the team to 20–25 people by the end of the year. Hughes called the area a good fit thanks to its high concentration of AI specialists. At the same time, he spoke out against the common UK practice of garden leave, under which a departing employee is barred from joining competitors or founding their own company for several months: the head explained that he himself faced such a restriction before founding Inherent together with other former Google DeepMind staff. ## 🔍 Fact-Check Verification - [Former Google DeepMind employees created Faraday, an AI agent that outperformed Anthropic and OpenAI on scientific tasks window-new](https://3dnews.ru/1147302/vihodtsi-iz-google-deepmind-sozdali-iiagenta-faraday-kotoriy-prevzoshyol-anthropic-i-openai-v-nauchnih-zadachah) - Базовый текст: Inherent, Faraday, 27 млрд параметров, $50 млн, Edward Hughes, RL, King's Cross, Garden leave. - [Inherent Says 27 Billion-Parameter AI Agent Beats Larger OpenAI, Anthropic Models In Scientific Research Test](https://www.tekedia.com/inherent-says-27-billion-parameter-ai-agent-beats-larger-openai-anthropic-models-in-scientific-research-test/) - Подтверждает 27 млрд параметров и сравнение с более крупными моделями; формулировка 'Inherent Says' указывает на самоотчётность. - [DeepMind alumni's AI agent beats OpenAI's GPT-5.5 in research tasks](https://www.newsbytesapp.com/news/science/deepmind-alumni-s-ai-agent-faraday-outperforms-openai-and-anthropic-models/story) - Подтверждает происхождение из DeepMind и сопоставление с GPT-5.5 в исследовательских задачах. - [Inherent's Faraday AI Outperforms OpenAI and Anthropic in Replicating Scientific Research](https://www.androguider.com/2026/08/inherents-faraday-ai-outperforms-openai.html) - Подтверждает задачу воспроизведения научных исследований; результаты не независимые. ## ❓ FAQ ### Q: What is Faraday and who developed it? **A:** Faraday is an AI agent from British startup Inherent, founded by former Google DeepMind staff. It can independently reproduce the findings of published scientific studies and verify their validity without any prior prompts. ### Q: What model does Faraday run on and how large is it? **A:** The system is built on the Qwen 3.6 model with 27 billion parameters — noticeably smaller than the compared Claude Opus 4.8 and GPT-5.5. ### Q: Is the claim of superiority over OpenAI and Anthropic reliable? **A:** These are the results of the startup's own internal tests, not of an independent benchmark. Moreover, Faraday uses OpenAI's GPT-5.5 Codex for programming, so a direct comparison with a "pure" GPT-5.5 is methodologically debatable. ### Q: What is the ultimate goal of the Inherent project? **A:** To create an AI scientist that not only checks existing conclusions but also acquires new scientific knowledge. Training is carried out using reinforcement learning.