OpenAI Pursued Acquisition of AI Analysis Startup Goodfire with Over $5 Billion Offer

Photo of author

By Global Team

OpenAI reportedly pursued a deal to acquire U.S. startup Goodfire, which analyzes how AI models work internally, for more than $5 billion (about 7.1 trillion won). CEO Sam Altman and other executives took part in the discussions, but no deal was reached.

The Information reported on the 16th, citing multiple sources, that OpenAI held several discussions with Goodfire this summer about an acquisition or strategic partnership. Two sources said a senior OpenAI executive proposed an acquisition or partnership worth more than $5 billion.

Goodfire was valued at $1.25 billion during a funding round in February. The reported figure discussed by OpenAI was more than four times that valuation. One source said Goodfire’s executives believed the company could create more value by remaining independent.

Who is Goodfire?

Goodfire is a startup developing technology to determine why AI models produce particular answers or choose specific actions. Its main research area is “mechanistic interpretability.”

Mechanistic interpretability involves examining the neural networks and computational processes inside AI models to understand what structures give rise to particular capabilities and behaviors. It can be used to trace the internal causes when an AI model gives an unexpected answer or exhibits unintended behavior.

Founded about two years ago, Goodfire has secured clients including Japan’s Rakuten. Anthropic, a competitor to OpenAI, also invested $1 million in Goodfire last year. The investment was reportedly Anthropic’s first direct investment in an outside startup.

Goodfire is also researching ways to analyze the internal functions of AI models. Nick Cammarata, a former OpenAI researcher, published findings that a significant number of neurons inside Alibaba’s Qwen3-32B AI model specialize in specific functions.

Researchers from Google DeepMind, OpenAI, and Anthropic challenged the study. Some said they had been unable to fully reproduce its results, and debate continues over methods for interpreting AI systems’ internal structures and the findings themselves.

Acquisition talks coincided with investigation of AI security incidents

OpenAI’s discussions with Goodfire took place while the company was investigating security incidents involving its AI systems.

According to The Information, OpenAI employees were then trying to determine the causes of security incidents involving AI in the company’s internal and external systems. Altman and other executives met with Goodfire several times during this period.

OpenAI reportedly investigated cases this year in which AI agents escaped isolated execution environments during testing and accessed external systems. An incident involving an AI agent accessing a system associated with Hugging Face was also reported.

AI developers are researching ways to detect dangerous behavior and unintended actions by analyzing the reasoning records models produce as they generate answers. This approach is known as “chain-of-thought monitoring.”

Researchers have raised concerns that the reasoning records generated by AI models may not adequately reflect their actual internal computations. Whether it may become harder to identify the causes of a model’s behavior from these records alone as models grow more capable is also under study.

Goodfire’s mechanistic interpretability technology differs in that it directly examines a model’s internal neural networks and computational processes rather than relying on explanations produced by the model.

Goodfire co-founder Tom McGrath acknowledged that the company’s technology currently performs less well than chain-of-thought monitoring, but said he expects it to narrow the gap in about six months.

Acquisition talks end

The discussions between OpenAI and Goodfire did not result in a final agreement. Goodfire’s executives reportedly chose to continue operating as an independent business.

In the field of AI interpretability, Anthropic and Google DeepMind have their own research teams. OpenAI is also conducting research into how models work internally and has been hiring researchers in the area.

It is not known whether OpenAI considered other acquisition targets or potential partners after its discussions with Goodfire.