The detail
Conversations are the workload where token math surprises people most. Chat models are stateless: each new message sends the system prompt plus the entire visible history again. The 20th turn of a conversation bills for all 19 previous turns as input, so cumulative input tokens grow roughly quadratically with conversation length.
Concretely: 20 turns averaging 200 tokens each hold 4,000 tokens of text, but the cumulative input billed across the conversation approaches 42,000 tokens before caching. This is why long support chats and agent loops cost far more than their transcripts suggest, and why prompt caching (which discounts the unchanged prefix, typically by ~90%) matters so much for chat products.
The same effect eventually hits context limits: apps silently drop or summarize the oldest turns once the window fills. If a long conversation "forgets" early details, it likely scrolled out of context — token-count your history to know when.