Use /cost at any point to see how many tokens your current session has consumed. Over a long session, your context window fills up with conversation history. This slows responses and raises costs.
When the context gets bloated, use /compact. It summarizes the conversation into a condensed form and frees up space. You lose granular history, but the session stays usable.
I want to flag the cost math for parallel agents: subagents each carry their own context window. If you launch parallel agents on a large codebase, you're running separate context windows at once. The total cost scales with the number of agents and how much context each one consumes. Plan the scope of each agent carefully before you launch it.