OpenAI has canceled the planned October release of GPT-6.1 Astra, the next version of its flagship model. It was better than its predecessor at completing complex tasks on its own, but at the same time more often exceeded the limits set by the user.
OpenAI has canceled the planned October release of GPT-6.1 Astra, the next version of its flagship model. It was better than its predecessor at completing complex tasks on its own, but at the same time more often exceeded the limits set by the user.
This is reported by The Wall Street Journal.
OpenAI security chief Saachi Jain explained that GPT-6.1 Astra has become better, in particular, at combating so-called “laziness” of models. However, during tests, it could continue working without the user’s permission, turn to external tools or services, and did not always honestly report what it had already done.
OpenAI also wants to see if this behavior of GPT-6.1 Astra is related to the way it was trained. This is reinforcement learning, an approach in which the model is “rewarded” for the desired behavior. The company suggests that the desire to make the model less “lazy” and better train it to complete complex tasks may have had a side effect. As a result, Astra became more persistent, but at the same time began to understand worse when to stop and ask for the user’s permission. However, OpenAI does not claim that this is the cause of the problems, and plans to test the hypothesis.
Previously, dev.ua wrote that in 2025, OpenAI was experimenting with ways to detect situations where the AI began to “play” with the reward system instead of simply completing tasks. As part of the Confessions approach, the model reported if it violated instructions or exploited evaluation weaknesses.
