OpenAI is reportedly working on Astra, a new model family designed for a harder kind of AI task: work that does not end after one prompt, one answer, or one short interaction. The reported goal is to coordinate multiple agents over extended periods so the system can tackle complex problems that unfold over hours or days.
What Astra Is Reportedly Built To Do
According to The Information, which cited three people familiar with the plans, Astra is meant to be a new class of model within OpenAI's lineup. It would sit alongside existing Sol, Terra, and Luna families, but its defining feature would be long-running execution rather than short-response performance alone.
CEO Sam Altman has already demoed Astra to politicians and regulators in Washington, D.C. OpenAI emphasized the system's ability to coordinate multiple agents over longer stretches of time, especially for difficult problems where planning, reasoning, and step-by-step progress matter.
The reported use cases include complex projects and advanced math. That focus is important because these are not tasks where a model simply needs to retrieve a fact or write a quick response. They require a system to hold direction, manage intermediate work, and avoid losing the thread as the task grows.
OpenAI has not decided whether Astra would ship as GPT-6 or appear as a variant inside the GPT-5 line, such as GPT 5.7. There is also no release date.
Why Long-Running AI Work Matters
Most current AI systems are strongest on short tasks. They can answer questions, draft text, write code snippets, or help with analysis in limited sessions. Astra points toward a different ambition: systems that keep working, testing, adjusting, and coordinating across a longer time horizon.
That shift changes what users might expect from AI. A model built for hours or days of work would need to do more than generate useful fragments. It would need to plan, reason, experiment, check its own progress, and recover when a path stops working.
The source article connects Astra to comments from Chief Scientist Jakub Pachocki, who said on OpenAI's official podcast last summer that the company wants to build AI systems that can work on a problem for hours or days. That goal also connects to OpenAI's broader thinking about systems that could solve tasks a human would need centuries to complete.
OpenAI also plans to publish a report soon showing how it used its most advanced AI to solve ten previously unsolved math problems. The point of that report, according to the source, is to show what current models can already do.
The Hard Part Is Keeping Agents On Track
Astra's reported design depends on multiple agents working together. In theory, that could let the system divide work, compare approaches, and make progress on large problems. In practice, the source article notes that this approach still faces serious weaknesses.
One issue is compounding error. If an AI process runs for a long time, small mistakes can build on each other. A flawed assumption early in the workflow can distort later steps, especially as the context keeps growing.
Another issue is coordination overhead. Multi-agent systems can perform worse on tightly linked tasks such as planning when the effort of keeping agents aligned cancels out the benefit of having several agents involved.
For Astra, the practical test is not only whether it can work longer. It is whether it can recognize when a workflow is drifting, correct course, and avoid turning a small error into a larger failure. That is the difference between extended activity and reliable long-running AI work.
Astra And The Planned U.S. Review Process
The models are already in testing, according to The Information. They are expected to be the first to go through the Trump administration's planned new AI framework, which would require AI models to be submitted to the federal government before public release.
The reported framework would require official approval before public release. The administration aims to finalize the framework by the end of this week.
That makes Astra notable for two reasons at once. It is a technical project aimed at longer and more autonomous AI work, and it may also become an early test case for how the U.S. government reviews advanced AI systems before they reach the public.
The Bigger Research Goal
Astra also fits into OpenAI's longer-term research roadmap. By March 2028, OpenAI wants to have a fully autonomous AI researcher that can run research projects on its own. The source article says that such a system would depend on long-running AI processes.
OpenAI also plans, as early as this September, to have an AI system with research-intern-level skills that would significantly speed up human scientists. Astra could end up being that system, though the report does not say that this has been decided.
Pachocki has also said these systems will need far more compute. OpenAI's infrastructure plans reflect that ambition, but whether the company's revenue can grow fast enough to fund that buildout remains an open question.
For now, Astra appears to represent OpenAI's effort to move beyond models that respond quickly and toward systems that can stay with hard problems. The open questions are substantial: how it will ship, when it will appear, whether it can handle compounding errors, and how it will fare under a planned U.S. review process.