OpenAI CEO Sam Altman broke the news again: after pushing back this summer’s “open model” release by a month, the company has now delayed it indefinitely. Originally slated to drop next week, the open‐weight model—whose underlying “weights” developers can download and run locally—won’t arrive any time soon. In a post on X (formerly Twitter), Altman cited the need for “additional safety tests” and a closer look at “high‑risk areas,” stressing that once the weights are out in the wild, there’s no recalling them.

When the open model was first announced, OpenAI set a timetable: release this summer, follow up on the heels of GPT‑5. But in mid‑June, Altman unexpectedly paused that schedule, noting the team had achieved something “unexpected and quite amazing” and wanted to ensure it held up across safety and reliability metrics. Now, less than a month later, he’s hit pause again. “We need time to run additional safety tests and review high‑risk areas. We are not yet sure how long it will take us,” he wrote, offering no new timeline.

Unlike closed models—where the provider keeps weights behind API walls—open models democratize AI, allowing anyone to dive into the code, tweak it, and deploy it on local machines. That transparency fuels innovation: researchers can audit biases, tinker with architectures, even spin up specialized forks. But it also poses risks: bad actors could repurpose the model for deepfakes, spam, or other malicious uses. OpenAI’s caution reflects lessons from past releases, when open‐source versions of large language models powered misinformation campaigns, phishing bots, and other nuisances.

While we trust the community will build great things with this model, once weights are out, they can’t be pulled back. This is new for us and we want to get it right.

This delay comes at a high‑profile moment. Investors and developers alike are itching for GPT‑5, the next iteration of OpenAI’s flagship model, and many expected the open version to arrive as a swan song before that launch. By postponing, OpenAI risks ceding headline space to rivals. DeepMind, Anthropic, and Elon Musk’s xAI are pouring billions into competing systems. An indefinitely delayed open model could raise doubts about OpenAI’s edge in a market where every week brings news of new benchmarks and breakthroughs.

In fact, on the very day Altman announced the delay, China’s Moonshot AI unveiled Kimi K2, a one‑trillion‑parameter open model that, according to company data, outperforms OpenAI’s GPT‑4.1 on several coding and reasoning benchmarks. Free to use, Kimi K2 scored 53.7% on LiveCodeBench—easily beating GPT‑4.1’s 44.7%—and posted a 97.4% accuracy on MATH‑500 compared to 92.4% for GPT‑4.1. That kind of performance (and price tag: zero) puts pressure on OpenAI to not only deliver but outdo the competition.

For the AI community, an open‑weight release is the ultimate playground. Startups could prototype niche applications without API fees. Academic labs could push forward transparency research. Even hobbyists would get under-the-hood access to state‑of‑the‑art reasoning capabilities. With the indefinite delay, that entire pipeline goes on hold. Developers who have been building integration plans around next week’s launch now face an open question: pivot to alternative open models, or wait on OpenAI’s safety sign‑off?.