Comment by panarky

It's cheap because it's a Flash model, far smaller and much less compute for inference, runs on TPUs instead of GPUs.