OpenAI Develops New 'Astra' Model Family Designed to Work on Problems for Hours or Days

OpenAI is working on a new model family with the tentative name 'Astra', intended to be significantly more capable for long-duration tasks than previous models. CEO Sam Altman demonstrated Astra to politicians and regulators in Washington D.C. this week, highlighting the ability to have multiple agents collaborate over extended periods to solve difficult problems, with potential applications in complex projects and advanced mathematics. According to The Information, Astra will form a new model class alongside existing families Sol, Terra, and Luna, and it is undecided whether it will appear as GPT-6 or as a variant like GPT 5.7. The models are already in testing and are expected to be the first to go through the planned new framework by the Trump administration, which requires AI models to be submitted to the federal government before public release. The report mentions the risk of error escalation in long-running workflows and the challenge of self-correction, which are current weaknesses of agentic systems. OpenAI also aims to release a report showing its most advanced AI solving ten previously unsolved mathematical problems, and has goals to create a fully autonomous AI researcher by March 2028 and an AI system with capabilities of a research intern by September. Chief Scientist Jakub Pachocki stated in the official podcast that OpenAI wants to build models that can plan, think, and experiment long-term, requiring significantly more computational power.