OpenAI is now paying a fifth more compute for safety checks on its riskiest work.
The company confirmed this week that it paused some frontier training runs for two weeks. It has also switched on a new monitoring system focused on its highest-risk projects. OpenAI says that system adds roughly 20 percent to the compute cost of whatever it covers. The company framed both moves as safety measures, though it gave few details on what prompted the pause or how it defines riskiest work.
A 20 percent compute tax is a concrete number, not a vague promise, and it shows OpenAI is willing to spend to slow itself down. That is notable for a company whose public posture has mostly been about speed, shipping models faster than rivals rather than throttling them. It also raises a question for the rest of the industry: does safety monitoring become a cost that only well-funded labs can afford to pay.
Whether that 20 percent buys real safety or just a compliance receipt is the part OpenAI has not shown yet.