It seems that with GPT-5.6 Sol-Ultra on fast mode you can burn through roughly 250 million tokens per hour per task?