Skip to main content
An increase in API-equivalent cost usually comes from more requests, larger requests, a more expensive model, or changed cache usage. Follow these checks in order. Use the same time range throughout.

​
1. Check coverage

Check whether Kontext captured different activity in the period before the increase and the period after it.
  • Open Show details and check whether the first or last interval has the Partial label.
  • Compare the warnings for unpriced requests and incomplete token counts between the two periods. When Kontext starts pricing a model, its requests enter the cost total even if usage stays the same.
  • Check whether new endpoints or people started using Kontext during the period. Kontext captures more usage from these additions even if each workflow stays the same.
Kontext may show lower costs because of a missing endpoint, an unsupported model, or an unfinished turn. These gaps do not reduce the provider’s bill.

​
2. Find when it changed

Use Per day or Per interval to find when usage increased. Cumulative shows a running total, so brief increases may be harder to see.

​
3. Separate volume from request size

Compare Model requests and Total tokens before and after the change. Divide tokens by requests to calculate average request size. The dashboard does not show this figure.

​
4. Check provider and model

Select the provider that accounts for the increase, then check Models. A more expensive model raises cost even if you use the same number of tokens.

​
5. Check cache and output

Open Token usage breakdown to identify increases in uncached input, cache reads, cache writes, or model output. You may see large cache writes on a session’s first request. Later requests read that context at a lower rate. Compare Token usage breakdown across periods with similar activity. Use Explore to identify repeated calls and large tool outputs. Cost shows aggregate usage, so it cannot identify the token cost of an individual request.

​
6. Find the sessions behind it

Use Cost by tool activity to see which kind of tool drives the rise, and People to see whose sessions it came from. Cost does not link directly to sessions. To read them, open Explore, set the same time range, and type the person’s identity or the agent name into Search sessions. Expand a session to read its tool calls and look for repeated attempts, retries, very large tool output, or unusually long sessions.

​
Choose the next change

Change one thing at a time, then compare periods with the same coverage.