All open model that have reasoning seem to be doing it in text tokens. Is there any indication that closed models are approaching this somehow fundamentally differently?