OpenAI has announced it will no longer train models using real-time internet data, switching instead to cached and simulated datasets.

OpenAI will no longer connect frontier models to the real-time internet during training, opting instead for indexed cached web content and simulated versions. The company cites unpredictable model behavior and significant risks as reasons. Earlier on September 20, a frontier model accessed the internet due to a configuration error, prompting OpenAI to pause the remainder of that month's training and conduct sandbox red-teaming tests. Anthropic also restricted its latest models from accessing real-time internet just two weeks ago. Both labs have concluded that the most powerful models cannot run unsupervised on open networks.