Comment by meandmycode

Also, while the cost of tokens is lower, let's argue it's cheap enough not to care. Reading 1m tokens surely isn't realistic for latency?