“The opening stages of OpenAI’s unraveling”: OpenAI slows model training — not everyone is buying the explanation

OpenAI says it has slowed part of its model development because its safety and monitoring systems weren’t keeping up. In a Tuesday blog post, the company said it has locked risky code away from the internet and from other internal systems, and that it is watching for dangerous behavior within 30 minutes. It also said the monitoring layer alone could add about 20% overhead to the inference compute it covers.

The company says the slowdown included a two-week pause in reinforcement learning for its latest models meant for deployment. Its largest planned frontier RL run is still on hold, with smaller training and evaluation rounds happening first. Sam Altman framed that as a straightforward trade-off: model progress is moving very fast, and OpenAI chose to slow down rather than let capabilities outrun alignment, security and monitoring.

But the wording leaves some room for interpretation. Altman later said OpenAI still expects to ship new models soon, and that the pause affects farther-out releases. At the same time, Time reported that a significant number of Astra workloads are still paused and being restored gradually, which suggests the slowdown may be broader than a single training run.

Alignment is the real buzzword here. OpenAI uses it repeatedly, and the idea is simple enough: a model can be powerful without being reliably steerable. According to reporting from Time, the pause was prompted by a collection of research observations showing different degrees of misalignment as capabilities advanced faster than researchers expected.

The trouble is that timing invites suspicion. On the same day, the Wall Street Journal reported that OpenAI’s operating losses widened by $3 billion to $12.3 billion in the second quarter, while its revenue grew 18% to $6.7 billion. CNBC also reported that CFO Sara Friar is eyeing a public-market entry in 2027, or sooner if the business keeps inflecting. That doesn’t prove the training pause was about money. It does make the safety sermon sound a little less pure.