Cursor Token Counter Over Counting by x10

Support Bug Reports
1

Describe the Bug

Cursor Token Counter Bug Report
Summary
Token counter displays inflated usage (~10x actual tokens) causing inaccurate tracking of API consumption.

Environment
OS: Windows 10 (Build 19045)
Cursor Version: [Current version]
Shell: Git Bash
Model: Claude Sonnet 4
Bug Details
Expected Behavior: Token counter should accurately reflect actual token usage Actual Behavior: Counter shows ~10x inflated token count

Reproduction Steps
Send simple prompt: “How is the weather in Sydney Australia” (~7 tokens)
Receive standard response (~197 tokens)
Observe token counter display
Evidence
Actual Usage: ~204 tokens total (7 prompt + 197 response)
Cursor Display: ~2000 tokens (2% of 200k limit)
Discrepancy: ~10x inflation
Calculation: 2% of 200k = 4000 tokens, but actual verified count is only 204
Impact
Users cannot accurately track token consumption
May cause unnecessary concern about usage limits
Affects cost/usage planning and budgeting
Could lead to premature usage restrictions

Steps to Reproduce

I don’t know

Operating System

Windows 10/11

Current Cursor Version (Menu → About Cursor → Copy)

Version: 1.4.4 (user setup)
VSCode Version: 1.99.3
Commit: dd05130a0bf8d5f1a5372e6e6bb7c4184786eaa0
Date: 2025-08-10T22:39:50.197Z
Electron: 34.5.1
Chromium: 132.0.6834.210
Node.js: 20.19.0
V8: 13.2.152.41-electron.0
OS: Windows_NT x64 10.0.19045

Does this stop you from using Cursor

No - Cursor works, but with this issue

2

Probably the system prompt. I don’t know details, but I think Cursor doesn’t just send our message straight to the LLM. It modifies it to add a system prompt adding more context, including any rules, etc.

3

This is a fair assumption, but at a x10 token penalty is completely unacceptable!!!

4

I think it would help you a bit to learn how it works — the whole tokens thing, which model you used, thinking or not,
and what the task was,
and whether tools were activated.

5

No sh*t Sherlock

6

Are you a programmer?
This is how you track down bugs.

7

each tool call uses whole chat history so 10 tool calls means 10x context? something like that, somebody from team described this earlier

8

Check the dev tools you’ll see Cursor unnecessarily adding tokens with no real purpose, just to burn through your quota faster. Classic trick to drain your tokens and make you spend more money.

9

I think it’s only 10x because in your test you sent a very short message, which was a lot smaller than the included system prompt.

If you sent a long message, then by comparison the system prompt would be much smaller, maybe a 10% increase instead of 10x.

10

It I have calculated the deltas between each response. While allowing for thinking model at times and non thinking in others the average is still 10x - not 10%