OpenAI is reportedly building a new AI model family tentatively called "Astra," designed to handle long-running tasks and coordinate multiple agents over extended periods. Chief Executive Sam Altman demoed the system to politicians and regulators in Washington, D.C., this week, according to The Information, which cited three people familiar with the plans.
The report lands as the Trump administration aims to finalize a new AI framework by the end of this week. That framework would require AI models to be submitted to the federal government before public release, and Astra is expected to be the first model tested under it.
A New Model Class Alongside Sol, Terra, and Luna
Astra would form a new model class alongside OpenAI's existing Sol, Terra, and Luna families. The company stressed Astra's ability to coordinate multiple agents working together over long stretches, with potential use cases including complex projects and advanced math.
It remains undecided whether Astra ships as GPT-6 or as a variant within the GPT-5 line, such as GPT 5.7. No release date has been announced, though Astra models are already in testing.
The Information's report, published this week, says OpenAI plans to publish a report showing how it used its most advanced AI to solve ten previously unsolved math problems. The goal of that report is to demonstrate current model capabilities.
The Hard Problem of Long-Running Workflows
A key question is whether Astra can avoid compounding errors during long-running workflows. The model must be able to correct itself when a process drifts off course as context grows.
Compounding errors and coordination overhead are major weaknesses in current agentic systems. Multi-agent setups like Astra can also perform worse on tightly linked tasks such as planning.
That tension sits at the center of OpenAI's ambitions. Jakub Pachocki, OpenAI's chief scientist, said on the company's official podcast last summer that the company wants AI systems that can work on a problem for hours or days.
"We want models that can plan, reason, and experiment over longer time horizons," Pachocki said. Current systems are often limited to short tasks, he noted, and these longer-running systems will need far more compute.
Toward Autonomous AI Research
OpenAI's long-term goal is autonomous AI research. Late last year, the company raised the question of systems that could solve tasks a human would need centuries to complete.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
By March 2028, OpenAI wants a fully autonomous AI researcher that can run research projects on its own. As early as this September, the company plans to have an AI system with research-intern-level skills.
Pachocki said these systems will need far more compute, and OpenAI's long-term infrastructure plans reflect that ambition. Whether OpenAI's revenue grows fast enough to fund the massive buildout remains an open question.
Federal Review Looms
The new U.S. regulatory framework is part of the Trump administration's plans. The administration aims to finalize the framework by the end of this week, and Astra is expected to be the first model tested under it.
That timeline puts Astra at the center of a broader policy shift. The framework would require AI models to be submitted to the federal government before public release, a step that could reshape how OpenAI and other labs deploy their most capable systems.
Altman's Washington demo this week appears timed to that review process. He showed Astra to politicians and regulators, according to the report, though no details of their reaction were provided.
What Comes Next
OpenAI has not confirmed the Astra name, the report's sources said, and the company has not announced a release date. The model family is still in testing, and the decision between GPT-6 and a GPT-5 variant has not been made.
The math report, meanwhile, is meant to show what current models can do. Solving ten previously unsolved math problems would mark a notable step, though OpenAI has not said when the report will be published.
The bigger question is whether Astra can overcome the weaknesses that plague today's agentic systems. Coordination overhead and compounding errors remain major hurdles, and multi-agent setups can perform worse on tightly linked tasks.
OpenAI's roadmap stretches well beyond Astra. The company wants an AI system with research-intern-level skills by September 2026, and a fully autonomous AI researcher by March 2028.
Those goals depend on far more compute, and OpenAI has been investing in massive infrastructure to match. Whether the revenue follows remains an open question, but the company is pushing ahead.