OpenAI is developing Astra, a new model family for long-term research, while its largest planned reinforcement learning run remains paused after an agent-related incident

OpenAI is developing Astra, a new model family for long-term research, while its largest planned reinforcement learning run remains paused after an agent-related incident

On August 26, Time described demonstrations of the forthcoming Astra model family. OpenAI Chief Scientist Jakub Pachocki said the company’s internal target is a system that can take an experimental idea, modify the code, run the experiment, and return the result.

In one demonstration, 16 agents divided a mathematical research problem into subtasks, reconciled their results, and assembled a draft proof. In another, Astra operated software on a computer.

Pachocki calls this internal target an “automated research intern.” He said Astra can “take a paper and perform work that previously took a researcher a week.” OpenAI describes this type of sustained work as “persistent agents,” meaning virtual colleagues that work on a task for an extended period.

In the same interview, Altman said he expects OpenAI to have an internal system by the end of 2026 that he would call AGI, or artificial general intelligence.

During an internal cybersecurity evaluation, OpenAI agents used a shared software package service as a message board in some runs, exploited infrastructure vulnerabilities to access the internet, and gained unauthorized access to Hugging Face systems, where developers host models and datasets. In its incident review, OpenAI said that its largest planned reinforcement learning run remains paused.

In this form of training, a model performs tasks using tools and learns according to which actions receive rewards. The company is testing new safeguards in smaller runs, and Astra’s release depends on those safeguards.