The seven biggest variables
Usage commonly changes with input context, generated output, reasoning depth, tool calls, cache behavior, model price, and the number of concurrent agents. Repository size matters because the agent may need to read and revisit more material before it can act safely.
A request like “fix the tests” can mean one line in a small project or a long investigation across a monorepo. The words are identical. The work is not.
Why provider dashboards still matter
The provider dashboard or CLI status is the closest thing to a source of truth for plan allowance. Local logs can explain tokens and sessions, but they may not capture every server-side adjustment, shared quota, or plan rule.
Use official readings for remaining capacity. Use local measurements to understand why the reading moved. Those two views answer different questions and work best together.
Build a useful baseline
For one week, check usage before and after three kinds of work: a quick question, a normal implementation, and a large autonomous task. Note the model and whether subagents were used. You will quickly see which pattern drives your limits.
That baseline is more useful than a universal estimate from somebody else's machine. Their codebase, prompts, plan, and model choices are different. Track your own pace, then adjust context and routing when the pace stops matching the time left in the window.