Cursor always turns 3 sentences into 60k+ tokens

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

I have only +132 tokens of context in my Cursor chat instance. When I submit a request to Agent or Plan (those are the only I’ve tried), I briefly see a loading state followed by an error about context overload. I have no files open. Only one Cursor window open. The only enabled model is Sonnet 4 via Anthropic API. My requests are only about 3 sentences; nothing complicated. But half the time I submit a request, I see an error message like the following, even if I wait a few minutes before clicking Resume or Try Again. Does Cursor have a log where I can see why it’s turning 3 sentences into 30k+ tokens constantly? I wish Cursor would alert me to this issue before spamming the hell out of Claude. My Claude logs show 60k+ input tokens in every request, with only about 400 output tokens.

{“error”:{“type”:“provider”,“reason”:“provider_error”,“message”:“Provider returned 429”,“retryable”:true,“provider”:{“status”:429,“body”:“{"type":"error","error":{"type":"rate_limit_error","message":"This request would exceed the rate limit for your organization of 30,000 input tokens per minute. For details, refer to: Rate limits - Claude Docs. You can see the response headers for current usage. Please reduce the prompt length or the maximum tokens requested, or try again later. You may also contact sales at Contact sales | Claude to discuss your options for a rate limit increase."},"request_id":"censored"}”}}}

Steps to Reproduce

Open Cursor > Open recent project > New chat > Select Sonnet 4 API model only > Attach a screenshot of the product and explain the UI bug, like too much space between the system tray and title bar, and ask for it to be fixed > Submit

Expected Behavior

Loading state followed by coding progress

Screenshots / Screen Recordings

Operating System

MacOS

Current Cursor Version (Menu → About Cursor → Copy)

Version: 2.3.35
VSCode Version: 1.105.1
Commit: cf8353edc265f5e46b798bfb276861d0bf3bf120
Date: 2026-01-13T07:39:18.564Z
Electron: 37.7.0
Chromium: 138.0.7204.251
Node.js: 22.20.0
V8: 13.8.258.32-electron.0
OS: Darwin arm64 25.2.0

For AI issues: which model did you use?

Sonnet 4 via API key; everything else disabled

For AI issues: add Request ID with privacy disabled

req_011CXAmmDqQspAdU9fKdDBoD

Additional Information

I’m not a developer. Only 1 month of experience with vibe coding. I just activated a new Cursor account and Anthropic key about 24 hours ago, but I’m continuing to work on a mobile app that I was working on using a different Cursor account. It’s a simple mobile app, with only a few days of work in total. Nothing complicated. The whole thing was built in Cursor, by myself.

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

… and here are my corresponding Claude logs. Notice how fast they grow with each (minor) request. I wonder if this is a dark monetization strategy by Cursor :face_with_monocle:

Note: most of these logs were from requests before I clicked New Chat in Cursor. The most recent 2-3 items in the log followed me closing all Cursor windows and clicking New Chat. In Cursor, the green/red context numbers at the top of the chat screen fell from 12k to <1k, but the Claude side didn’t see any reduction at all. Creepy. Next I’ll try restarting the computer and starting fresh as much as possible again.

I restarted the computer and tried again. I started with a new chat in Cursor, in the same project. I sent only 2 simple requests to Agent in Cursor, but the Claude API logs show 20 requests, with input tokens totaling 642k since the reboot. Every request in the attached screenshot came after restarting the computer, launching Cursor, and sending 2 short requests, the first of which contained a screenshot of the app.

Hey @vibez!

132 tokens is just your input. Cursor sends a lot more behind the scenes.

For example:

  • The System prompt (thousands of tokens of agent instructions)
  • Your project structure
  • Any .cursor/rules/ files in your project
  • User rules from Settings > Cursor Settings > Rules
  • Your screenshot. Images can easily add 10-20k+ tokens depending on resolution

Each request includes the new context generated from your previous request.

Any particular reason you’re using your own Anthropic API key instead of Cursor’s included usage? With BYOK you’re hitting Anthropic’s personal tier rate limits (30k tokens/min), which is pretty restrictive.

Hi, thanks for the reply.

Cursor’s dashboard doesn’t provide enough information to rely on it alone:

  • The Spending page has no information whatsoever. The only thing that appears on that page is a switch to turn on unlimited spending.

  • The Usage page shows per-request rows but no aggregate dollar or token totals, by hour, day, week, month, or lifetime.

  • There’s no way to tell how fast credits are being consumed or whether a spike came from a single action or background behavior. I don’t know if my current pattern of usage will eat through my $20 of credits in a few days or a few months.

That Claude Console shows a “total token cost,” as well as summaries and charts about token usage. That’s how I was able to diagnose this unexpected burn rate, where every Cursor interaction was triggering many streaming requests, each with tens of thousands of input tokens.

Cursor needs to improve its usage information. In addition to adding the Claude features to your web console, when Cursor completes a task, I’d like to see a dolarized cost estimation inside the app about how much that task just cost. This way I could learn how expensive it is to attach screenshots to my question, or how much Agent mode costs vs. Ask or Debug.

In addition, Cursor should provide cost statistics per workspace. That would allow me to bill each client for my token costs associated with their project. The data is certainly available to Cursor, so why not provide it to your customers?

Appreciate your attention,

Jesse