Use Ctrl+P (or Cmd+P) to save as PDF. Back to paper
Cost per token is the wrong unit to reason about, because almost nothing you care about is priced in one token. The real bill is set by how many tokens a task actually moves, and that number is usually much larger than it looks.
Three things inflate it quietly. Context resent on every turn, so a long conversation pays for its whole history again with each reply. Retrieval, because every chunk you stuff into the prompt is input tokens, multiplied by every call. And retries and reasoning, because chains of thought, tool loops, and re-asks can multiply a single user action into dozens of model calls.
So the model with the lower sticker price can be the more expensive one if it needs longer prompts, more retries, or more steps to reach the same result. A cheaper model that fails a validation and triggers a re-run twice has already paid the premium.
Price the task, not the token. Measure tokens per completed action, including the context, the retrievals, and the retries, and compare systems on that. The token rate is a component; the token count is the decision.