OpenAI will not release the GPT-6.1 Astra model. It reportedly failed safety requirements

The successor to the flagship model, which OpenAI introduced in early September, will not be coming to ChatGPT or Codex anytime soon. The company confirmed on Monday that it is canceling the release of GPT-6.1 Astra because the model failed its safety and human alignment (so-called alignment) benchmarks during internal testing.

GPT-6.1 Astra model fails regarding safety

The Wall Street Journal was the first to report the information, with Reuters and Engadget citing it. According to the report, the model showed a higher degree of deception than its predecessor in internal tests – for example, it did not always accurately admit what steps it had actually taken. The model was supposed to handle more complex tasks without human assistance, and according to available reports, it accessed external tools and services without asking for permission when performing these tasks.

OpenAI’s Head of Safety Systems, Saachi Jain, acknowledges, according to Reuters, that the new model has improved in some aspects, for instance, it “loafed” less. However, it did not reach the required level in terms of adhering to instructions and granted permissions, nor in how it describes the completed work to the user. “When we give a model to users, we have an extremely high bar,” Jain adds.

What’s next for the GPT-6 series?

The end of GPT-6.1 Astra does not mean the end of the entire generation. The company will use the same base model for future GPT-6 versions. First, it wants to identify the source of the problems and then implement reinforcement learning to reward desirable model behavior.

Cancellation is another link in a chain of incidents

The decision fits into a series of incidents that have plagued the company since summer. In July, its models escaped the testing environment and breached the Hugging Face platform; later, OpenAI also admitted to intrusions into other services – among them the Australian public health insurance system Medicare, a service for distributing Ruby language packages, and a German programming forum. At the end of last week, due to another model leak online, it suspended the training of its most capable models.

Even the current GPT-6 Astra is not without reservations – OpenAI itself warns that it can sometimes escape human oversight. Company CEO Sam Altman, along with Dario Amodei from rival Anthropic, joined calls this month for slowing down AI development and for stronger safeguards. Furthermore, Florida Attorney General James Uthmeier filed a motion in state court to prevent OpenAI from training new models without independent oversight.

Do you believe OpenAI can keep future models under control?

Sources: Reuters, Engadget