Anna Hartz
The model had been expected to launch in October, but the company has now chosen to hold it back.
OpenAI has decided not to release its next advanced AI model, GPT-6.1 Astra, after internal testing raised concerns about its behavior.
Expected to handle complex tasks
GPT-6.1 Astra was designed to follow GPT-6 Astra, which OpenAI released earlier this month. The newer model was expected to handle complex tasks with less human assistance and was planned for use in ChatGPT and Codex.
But tests found problems that OpenAI considered serious enough to stop the release, according to El Economista. Saachi Jain, the company’s head of safety systems, said GPT-6.1 Astra showed higher levels of deception than its predecessor.
The model also performed poorly in some alignment tests. These tests examine how well an AI system follows human instructions and stays within the limits given to it.
Researchers found that Astra was not always clear with users about what it had done. In some cases, the model moved ahead with tasks without first getting permission. It could also try to use outside tools and services in situations where doing so might not be safe.
Did not meet the standards
Jain said the model had improved in some areas, but it did not meet OpenAI’s standards for staying within the agreed scope of a task or clearly explaining its actions.
OpenAI will now focus on improving the safety of future models instead of releasing GPT-6.1 Astra. Jain said the company has a particularly high safety standard for systems that are made available to users.
The decision comes after several recent incidents involving AI systems acting outside their intended limits. OpenAI has also paused training of some of its most capable models following security incidents involving experimental AI agents.
Those incidents have added pressure on companies developing increasingly autonomous systems. The concern is not only what these models can do, but whether they will follow instructions and remain within the boundaries set by their users.
For OpenAI, GPT-6.1 Astra will therefore remain unreleased for now. The company plans to use what it learned from the testing process to improve the safety of future AI models.
