OpenAI explored acquiring artificial intelligence startup Goodfire for more than $5 billion as part of discussions aimed at improving its understanding of how AI models behave and identifying potential security risks, according to a report by The Information.
The discussions took place during the summer of 2026 and involved OpenAI CEO Sam Altman and other senior executives, who held multiple conversations with the San Francisco-based startup. A senior OpenAI executive reportedly raised the possibility of an acquisition or an alternative commercial partnership.
Goodfire, which is backed by Anthropic, specialises in mechanistic interpretability, a field of AI research focused on understanding the internal computations that influence how machine learning models produce outputs and make decisions.
The proposed acquisition price would have exceeded Goodfire's $1.25 billion valuation announced in February 2026 by more than four times. However, the discussions did not result in an acquisition agreement.
According to the report, one reason the talks lost momentum was Goodfire's belief that remaining independent could create greater long-term value by allowing it to provide technology to multiple AI developers.
The discussions followed security incidents involving AI systems, including a reported breach affecting Hugging Face. These incidents have raised questions about how developers can monitor increasingly autonomous AI models and prevent unintended or unauthorised actions.
Goodfire's technology examines mathematical activity inside AI models rather than relying solely on their written reasoning processes, commonly described as chains of thought.
Traditional monitoring approaches often analyse these reasoning traces alongside a model's observable actions to identify potentially harmful behaviour. However, researchers have raised concerns that such methods may become less reliable as AI systems evolve.
Goodfire acknowledges that mechanistic interpretability has not yet reached the maturity of established monitoring techniques. Its co-founder and chief scientist, Tom McGrath, expects the approach to become more competitive and less expensive as the technology develops.
The company offers annual software licences and usage-based access to Silico, a research platform that uses AI agents to conduct experiments on model behaviour. Its customers have included Japanese technology company Rakuten and AI neuroscience company Prima Mente.
Goodfire has generated revenue, although its financial performance has not been publicly disclosed.
Anthropic invested $1 million in Goodfire in 2025, marking its first direct investment in another startup. The Claude developer has also conducted extensive research into mechanistic interpretability as part of its AI safety work.
OpenAI has separately been recruiting interpretability researchers and incorporating related techniques into evaluations designed to identify potentially unsafe model behaviour.
The reported discussions indicate growing commercial interest in technologies that could help AI developers understand and monitor increasingly complex systems.
Neither a completed acquisition nor an agreed purchase price was reported, and the discussions should not be treated as a formal transaction.
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.