← All writing

Checking what my AI usage numbers mean

An unexpected quota reset led me to examine pricing fallbacks, source freshness and the difference between estimates and spend.

An unexpected quota reset sent me back into my usage tracking on October 23. I wanted to know whether the numbers were accurate enough to help me use the subscriptions I already had.

The audit found a concrete pricing problem. An explicit cache-token price of zero was being treated as missing, which caused the code to use a fallback price. The assistant reported fixing that behavior. A valid zero needed to stay zero.

The affected total was an estimate of what the recorded usage would cost at API prices. That distinction is easy to lose when a dashboard shows a dollar amount. An API-equivalent estimate can help describe usage, but it doesn’t tell me what was charged to my subscription.

I also asked why the tracker needed several data sources when T3, the agent application I was using, already collected usage. The explanation separated the information each source supplied. Quota snapshots answered a different question from local transcript token history. Rate-limit telemetry added another view, with its own timestamps.

That raised a practical requirement for the display: I needed to know what a number measured and how current it was. If two sources described the same metric, their timestamps mattered. If they described different metrics, combining them into one total could obscure what either one meant.

The reset itself needed care as well. A reset deadline passing didn’t confirm that the provider had refilled the quota. Treating the schedule as an observed balance would create certainty the data didn’t support.

The fix was reported in the conversation, but tests weren’t run under the policy active during that work. I can describe the bug and the reported code change. I can’t use that record alone to claim the whole tracker was verified.

For future comparisons, I want to keep actual charges separate from estimated replacement cost. Subscription quota has its own use: helping me decide what capacity remains. Likewise, elapsed agent runtime can’t tell me how much active work I did or how much time an assistant saved. Those questions need different evidence, even if a single dashboard eventually presents them together.

Written by . Follow via RSS ↗