OpenAI moves its safety review ahead of the training runs it expects to raise capability

The change is in sequence. Responsible Scaling Policies and Preparedness Frameworks were aimed mainly at deploying completed models, Altman wrote ^[4]. The new safety case is written before a frontier reinforcement-learning run that OpenAI expects will significantly increase capability ^[2]. Whoever reads that document is reasoning about a model that does not exist yet, on a trigger the company sets for itself.

OpenAI has not named anyone outside the company to read one. The company pledged on Saturday to give independent evaluators employee-like access, matching Anthropic's commitment, and Altman said it would share more soon ^[9]. The post does not name the evaluators or say when their access begins, which runs count, who reviews the safety cases, whether any outside party sees them, or whether a run has been delayed ^[13].

Altman named the cost himself. "When we talk about 'pacing', we do not mean 'stopping'," he wrote. "Progress has been rapid and will continue to be. But it should be slower than it otherwise could be; interventions like safety cases and monitoring have significant costs." ^[10] He also wrote that "no amount of American competitive pressure should justify recklessness, or let capabilities get ahead of alignment and monitoring" ^[11].

Amodei's route needs Washington and Altman's does not. Dario Amodei asked for a narrow antitrust waiver so rival labs could discuss common safety measures, writing that it was needed "for antitrust reasons" ^[6]. Coordinated restraint is the part that needs legal cover, and whether enforcers would let companies slow development together is unclear ^[7]. Acting alone costs OpenAI money and leaves every competitor free. Altman is asking other companies to propose their own methods while the industry works toward common standards ^[12], and he called for shared standards on misalignment, monitoring and safety ^[8].

The hardest objection to the coordinated version came from the journalist Brian Merchant, who wrote that proposals similar to Amodei's "would likely only wind up serving Anthropic and OpenAI; it's what regulatory capture looks like in action" ^[14]. That charge lands on a waiver more than on a unilateral cost. A rule two labs write and everyone obeys raises rivals' costs; a document OpenAI writes for itself raises OpenAI's, and a rival is free to ignore it.

The federal route is slower. Trump dismissed the calls on Sunday, saying "We're leading China in AI. We're the most sophisticated country in the world, and frankly, I want to keep it that way because whoever wins AI wins" ^[17]. China's Foreign Ministry called the executives' comments "fearmongering" on Monday ^[18]. Altman said government help would be needed for international coordination but not for the first domestic steps: "But first we should do what we can ourselves" ^[19].

Investors treated the weekend as a cost. Implicator.ai reported AI-linked shares falling Monday as investors reacted to the statements ^[20]. SoftBank Group fell 13 percent in Tokyo on Monday, Sept. 14 ^[15], and Samsung Electronics and SK Hynix each lost more than 4 percent in Seoul that day ^[16]. SoftBank's was the largest of the three declines, at under 3.3 times the Korean drops ^[21].

There is one way to see whether the new gate binds: a run disclosed as delayed, altered or abandoned because of what a safety case said. Until then, OpenAI writes the safety case and OpenAI reviews it. Altman set the bar for everyone: "Every frontier lab must deliver on this, and there is no reason any of us should come to work if we cannot," he wrote ^[5].