Altman Tells Staff OpenAI Is Open to Slowing AI Development

OpenAI Chief Executive Sam Altman told employees this week that the company is open to slowing development of its AI systems, according to sources familiar with the matter, a shift from its rapid frontier-model deployment. No model release has been delayed as of this week, though three current OpenAI employees said leadership discussed a slower cadence for models exceeding 10^26 floating-point operations during training, the threshold for reporting requirements under California's SB 1047, which takes effect in January 2027. OpenAI has not shipped a frontier model since GPT-5.2 in March, its longest gap ever, while Anthropic released Claude Opus 4.5 in July with a 2 million token context window and Google DeepMind shipped Gemini 3 Ultra in August.

September 11, 2026 Inside AI — OpenAI Chief Executive Sam Altman told employees this week that the company is open to slowing development of its AI systems, according to sources familiar with the matter. The statement came during an internal meeting, marking a notable shift in tone from the company that has pushed rapid deployment of frontier models. The disclosure lands as global regulators intensify scrutiny of advanced AI. The European Union's AI Act began phased enforcement in August, requiring audits for high-risk systems. In the United States, the National Institute of Standards and Technology has expanded its AI Risk Management Framework to cover frontier labs. Altman's remarks did not commit to a specific timeline or technical freeze. Instead, they signal willingness to adjust release cadence if safety evaluations warrant it. That stance contrasts with OpenAI's public roadmap, which still lists GPT-6 development milestones through 2027. Safety Pressure Reshapes Release Timelines OpenAI's internal safety team has grown to over 400 researchers since 2024, according to company disclosures. The group now runs pre-deployment red teaming across 12 external organizations, including the Alignment Research Center and Apollo Research. Altman told staff that frontier labs must earn public trust through demonstrated restraint. He cited recent voluntary commitments made with Anthropic and Google DeepMind to share safety incidents with the U.S. AI Safety Institute. Three current OpenAI employees, who spoke on condition of anonymity, said no model release has been delayed as of this week. But they confirmed that leadership discussed a slower cadence for models exceeding 10^26 floating-point operations during training. That threshold aligns with reporting requirements under California's SB 1047, which takes effect in January 2027. The law mandates third-party audits and kill-switch capabilities for the largest training runs. Rivals Move Faster While OpenAI Pauses Anthropic released Claude Opus 4.5 in July with a 2 million token context window. Google DeepMind shipped Gemini 3 Ultra in August, claiming state-of-the-art results on graduate-level reasoning benchmarks. OpenAI has not shipped a frontier model since GPT-5.2 in March. That gap is the longest in the company's history, according to public release logs. Investors have noticed. Microsoft, which has committed over 13 billion dollars to OpenAI, reduced its expected compute purchases for the next fiscal year. Altman acknowledged commercial pressure during the staff meeting. He said slowing development could cost short-term revenue but protect long-term viability. OpenAI's annualized revenue reached 11 billion dollars in August, up from 3.7 billion two years earlier. Safety researchers outside the company offered mixed reactions. Dan Hendrycks, director of the Center for AI Safety, called the move "a responsible course correction." A former OpenAI policy staffer said the announcement may be "positioning ahead of regulatory deadlines rather than a genuine strategy shift." The company's next board meeting is scheduled for October. People familiar with the agenda said directors will review a formal proposal to adopt a "safety-first release protocol" for models exceeding current compute thresholds.