Models can complete longer and longer tasks, with the doubling time for task length reaching approximately seven months in recent studies, suggesting exponential improvement in sustained reasoning capability.