Frontier AI labs are shifting focus from faster answers to longer-running AI agents that work for hours, days, weeks, or months. OpenAI previewed a model family called Astra, built for long-running tasks, to policymakers in Washington DC this week. According to The Information, OpenAI touted Astra's ability to have multiple agents work together over long periods to solve hard problems, such as advanced math. Anthropic is describing runs measured in weeks with its newest models in the Mythos/Fable/Opus 5 families, focusing on AI Coding, Enterprise, and professional productivity. OpenAI published results on ten problems in mathematics and theoretical computer science with no progress for at least a decade, achieved by an internal version of Astra. The total number of tokens needed to find solutions would cost roughly $2,000 at Sol API rates. Boris Cherny, who leads Anthropic's Claude Code, described Opus 5 in a Y Combinator interview published July 27.