OpenAI is building a new model family, currently called Astra, meant to handle tasks that take hours or days instead of seconds. Sam Altman demoed it privately to US senators and regulators in Washington this week, according to reporting from The Information. The pitch: several AI agents split a hard problem, each works on a piece, then they combine results, with advanced math and large research projects given as examples.
This is not OpenAI's first attempt at organizing models into families. Just months ago, the company released Sol, Terra, and Luna as tiers of GPT-5.6, with Sol as the most powerful, Terra as a cheaper middle option, and Luna built for speed. Astra would sit above or alongside those, though OpenAI has not decided if it ships as GPT-6 or as a smaller update inside the GPT-5 line.
The more interesting detail is who gets to see it first. Astra is expected to be the first OpenAI model to go through a new federal review process, created by an executive order President Trump signed in June. Under that order, companies can voluntarily submit their most advanced models to the government for a security check before wider release. A final version of that framework was expected within days of the Astra demo. This is not mandatory law, it is a voluntary arrangement, but it sets a real precedent: the most capable AI systems now get a government look before the public does.
Before taking OpenAI's claims at face value, it is worth remembering its track record. Last year, the company said an internal model had solved ten previously unsolved math problems. After mathematicians reviewed the work, one of those claims had to be withdrawn. That does not mean the new report on math problem-solving is wrong, but it means the numbers deserve scrutiny before anyone treats them as settled fact.
The same caution applies to multi-agent systems generally. Coordinating several AI agents sounds efficient, but research on these setups shows a different story: production systems built this way fail a large share of the time, often because agents misread their role, duplicate work, or skip checking each other. Errors that would be minor in a single AI system can snowball across a long, multi-day task.
None of this changes the direction OpenAI is heading. The company has said it wants an AI system with research-intern-level skills soon, and a fully autonomous AI researcher by March 2028 that can run its own experiments without a human in the loop. Astra looks like an early step toward that. The catch is money: OpenAI's revenue is a fraction of the hundreds of billions of dollars in compute spending it has committed to through 2030. Whether that gap closes before the ambitions do is the real question hanging over this entire effort.