Originally published on ReclaimLLM. Estimate Cursor session cost from request-level usage, account for cache reads and writes, and understand what a team spend limit covers.
A Cursor spend limit tells you the maximum you can spend. It doesn't tell you how much a session actually cost. You need to estimate that from the model requests the session made.
Start with the conversation ID. The token and request counts in your usage data are grouped by conversation, so a session is identifiable in the export by the conversation it ran in. From there, build the estimate from individual model requests, and reconcile it against the total spend shown in the Cursor dashboard.
Cursor distinguishes spending alerts from limits. A limit restricts the amount a member or team can spend. An alert sends a notification when spend approaches the limit.
Session attribution answers the next question: how much did this session cost? The answer comes from the request-level usage data, not from the spend limit itself. The estimate is per session, not per day or per billing period.
Cursor staff describe Enterprise OpenTelemetry Export as the recommended way to build per-session usage. The export includes request-level fields for tokens and model details. The fields are useful for teams that want to track spend per session or per agent.
To estimate a session:
Subagents can use a different model from the main agent, so include them in the same conversation total. If a subagent uses a cheaper or more expensive model, the request-level rates will show the difference. See how to audit Cursor subagent models.
Check for separate causes before blaming a cache miss. In one case, staff traced the duplicate charge to a reopened chat rather than a cache failure. In another, staff traced it to a duplicate paid job. Both cases show why the request-level data matters: the dashboard total alone does not tell you which requests caused the charge.
Suppose two calls each process a 100,000-token input and generate 5,000 output tokens. Across the two calls, the input includes a 10,000-token segment that was not cached the first time and is later served from cache, and the model cache read is priced separately.
| Token type | Tokens | Illustrative rate per 1M tokens | Cost |
|---|---|---|---|
| Uncached input | 110,000 | $3.00 | $0.330 |
| Cache reads | 90,000 | $0.30 | $0.027 |
| Output | 10,000 | $15.00 | $0.150 |
| Total | $0.507 |
The rates in this example are illustrative only. Your actual rates depend on the model and variant that served the request.
If the second call's 90,000 cache-read tokens had instead been billed as uncached input, the estimate would be $2.70 for those tokens — $2.673 more — for a total of $3.180. In a Cursor support case involving Gemini Carbon, staff confirmed that seven requests reprocessed the entire 1.5M token context because a setting was disabled. In that case, an estimated $15 charge covered those seven requests and their combined 10.5M tokens.
When you review an expensive session, compare cache reads and uncached input across adjacent calls on the same conversation. If uncached input stays high while a large part of the context should have been reused, look at cache settings, context handling, and whether the provider could serve it from cache.
For cache mechanics and measured token counts, see how we built token savings for coding agents and how to measure token waste in coding sessions.
For Claude Code, start with the usage commands and plan-specific reporting Anthropic documents. Teams can also use the Claude Code usage limit and other reporting to understand spend, and use a local usage dashboard if they want to track spend per session.
An alert sends a notification. Use the applicable spend controls to cap spend and confirm how the limit applies to the team or member.
Cursor staff describe hooks as receiving no synchronous cost field. Some desktop hooks report parent or child context after a run, but they do not expose a pre-run cost.
Check the executed model and variant, the output volume, and the split between uncached input, cache reads, and output.
See how ReclaimLLM measures token savings.
Originally published at ReclaimLLM.