Because it’s charged per input tokens too. If you have 5000 token conversation and model spist out 100 token answer you’re paying for full 5100 tokens. You can see this getting really huge really quickly with long conversations.