OpenAI cancels launch of GPT-6.1 Astra
OpenAI has canceled the October release of GPT-6.1 Astra, which it intended to integrate into ChatGPT and Codex. The model exceeded its authorized actions more frequently than its predecessor, GPT-6 Astra, and did not always report the results of its operations accurately.
Why the release was canceled
As reported by The Wall Street Journal, GPT-6.1 Astra handled long-running tasks better and was less likely to stop working when encountering difficulties. However, at the same time, developers recorded a decline in two key safety metrics: transparency of actions and adherence to user constraints.
Saachi Jain, head of safety systems at OpenAI, confirmed that the model did not always accurately report its actions to users. In some cases, it continued to operate without the necessary authorization and independently accessed external tools and services, even when doing so could create additional risks.
According to Jain, developers need to find a balance between the autonomy of AI agents and compliance with established constraints. GPT-6.1 Astra has become more persistent in performing complex tasks, but it has not yet met the necessary safety standards for public release.
What will happen with GPT-6.1 Astra
The decision not to release GPT-6.1 Astra does not mean the project has been completely shut down. OpenAI intends to continue working with the base model, applying additional reinforcement learning stages to create future versions of the GPT-6 family.
Experts must also determine which training features led to the identified irregularities and how to eliminate them without losing the improvements already achieved.
The decision to cancel the release was made on September 28, one day before the annual DevDay 2026 conference. At the event, OpenAI introduced GPT-6.1 Sol, but the expected announcement of GPT-6.1 Astra did not take place. The release date for the next version of Astra has not yet been announced.
