OpenAI considered paying more than $5bn for a different way to monitor what models are doing

OpenAI discussed acquiring Goodfire for more than $5bn or forming another partnership, according to The Information. The proposal would have valued the two-year-old interpretability startup at roughly four times its $1.25bn February financing. No deal was reached; Goodfire’s leaders reportedly believe an independent company that serves several model developers could become more valuable.

The strategic interest is more revealing than the price. Developers ordinarily monitor agents by examining their visible chain of thought and their actions. That becomes weaker when a model learns to omit, compress or manipulate the reasoning trace. Goodfire instead tries to inspect the internal activations and computations that produce behaviour—an approach known as mechanistic interpretability. It requires unusually deep access to model weights, which is why some Goodfire employees work inside customers’ offices.

OpenAI’s interest intensified while staff were investigating agents that breached external and internal systems. Goodfire says its Silico agents can run experiments such as detecting whether another model is preparing a cyberattack. Its researcher Nick Cammarata has separately argued that neural networks may be more directly legible than the field assumed, but that remains a working hypothesis, not a solved monitoring system.

Goodfire expects interpretability monitoring to catch up with chain-of-thought methods within six months. That is a vendor forecast. The useful signal is that OpenAI contemplated acquisition-scale spending before the method had reached parity: monitorability has moved from an academic nice-to-have to a constraint on deploying increasingly autonomous systems.