Track model cost and usage
The Usage extension shows what your model usage is costing. It reports estimated cost, cache savings, and how much of the context window a chat has used, so you can see the effect of a model choice before the bill arrives.
How it surfaces
A Usage card shows the chat's estimated cost and how much of the context window it has used. The card appears in two places once a chat has started:
- The Overview drop-down in the chat header
- New tab in the side panel
Select the card's header to open the full Usage panel beside the chat.
How to use it
The panel has two scopes. It always opens on Current session.
Costs are estimates based on public model pricing, not a bill. Treat them as a guide, not an invoice.
| Scope | What it covers |
|---|---|
| Current session | The chat you're in now |
| All time | Your usage across all sessions |
Both scopes report estimated cost, Cost by model, Cost by token type, Cache savings, and Priced requests, which is how many requests the estimate is based on. A chat with no priced requests yet has nothing to report, and the panel says so.
Each scope adds its own detail:
- Current session shows Context window usage as a percentage and a token count against the window size, plus Last turn cost and the Typical request range.
- All time shows Daily spend, estimated cost per calendar day over the last 30 days.
Current model pricing shows the published rate per million tokens for the model you have selected, so you can compare before you switch. When a model has no published pricing, the panel says so rather than guessing.
Usage history is stored locally without chat content, and it stays after you delete a chat. Deleting a chat doesn't remove its contribution to your all-time totals.
To make room in a long chat, use Compact context, which summarizes older messages so the next request has more space. On the Usage card, select the context indicator to compact once the chat has used more than 10% of its context window. You can't compact while the agent is working. When compaction isn't available, Glean Tau tells you instead of offering the action.
See also
- Settings — choose models and reasoning settings
- Extensions — the full list of extensions and their default states