Former Google DeepMind researcher and AlphaGo co-creator Thore Graepel has launched Metis Reasoning, a London-based artificial intelligence company developing reasoning systems designed to operate with lower computing requirements than large frontier AI models.
The company, launched on October 20, 2026, is pursuing an alternative approach to AI development, focusing on computational efficiency, verifiability and reliability rather than increasing model sizes and training budgets.
Graepel, who left Google DeepMind earlier this year, has established Metis alongside technology entrepreneur Piers Ingram, robotics specialist Michael Warren and AI researcher David Berman.
The founding team brings experience across machine learning, robotics, autonomous systems, biotechnology and AI advisory services. Metis has assembled a 12-member team comprising researchers, mathematicians, robotics experts and technology entrepreneurs.
The company's research focuses on AI reasoning, particularly how systems evaluate information and work through complex problems. Metis believes its approach could support applications in engineering, robotics and other sectors where the ability to verify AI-generated outputs is important.
According to the company, existing large language models can perform tasks involving language understanding, instruction-following and multistep reasoning. However, Metis argues that limitations in the reliability and traceability of their reasoning processes can restrict adoption in applications requiring greater accountability.
Ingram said these concerns become more significant when AI systems are used in fields such as science, engineering, medicine and law, where decisions may require independent verification.
Metis is drawing on architectural concepts associated with AlphaGo, the Google DeepMind programme that demonstrated advanced decision-making capabilities in the board game Go.
The company maintains that similar principles could support the development of AI systems with more verifiable reasoning processes while requiring fewer computational resources.
Ingram described the proposed architecture as substantially more compute-efficient than existing frontier models.
“We are very much not going in this direction of ever larger models and ever larger training budgets,” he said.
He also claimed that the company had achieved encouraging results on advanced reasoning problems at a fraction of the computing cost associated with frontier models. Independent benchmark results were not disclosed in the report.
Metis is understood to be seeking funding worth tens of millions of dollars, although Ingram declined to confirm fundraising discussions. He said the company had secured sufficient funding for its operations to date.
The company has not announced a timeline for the public release of its first AI model.
Metis enters a growing London AI research ecosystem that includes Ineffable Intelligence, founded by former DeepMind researcher David Silver, and Recursive Superintelligence, co-founded by computer scientist Richard Socher.
The launch comes amid continuing investment in alternative AI architectures, as developers explore approaches to improving reasoning capabilities without relying exclusively on larger models and increased computing expenditure.
Disclaimer: This article may include information derived from interviews, press releases, public statements, research, company communications and other publicly available or third-party sources. Such material may be summarised, paraphrased or contextualised for journalistic and editorial purposes. All rights in third-party content remain with their respective owners.