Cursor sending almost entire project on every request, consuming all tokens instantly

[A_A6 (A A)]

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

Example request ID: 41c14510-3380-4862-b4dd-b2efb343adde

GPT-5.5-high - One question, about a bug, with a stacktrace … “worked for 2 mins 6 secs” … and consumed 227,500 tokens

I have - for months - asked similar questions on this codebase, and typically it would consume 30,000 tokens, up to (at most) 75,000 tokens.

NOTE: if it were just one conversation I wouldn’t call it a bug. In the last 24 hours its been every single conversation - dozens of them. This same issue is occurring across the project, with new sessions, any conversation.

Something is badly broken!

Steps to Reproduce

Any new conversation that asks a question about the codebase … appears to send the entire codebase (or close to it).

Operating System

Windows 10/11

Version Information

Version: 3.10.17 (user setup)
VS Code Extension API: 1.125.0
Commit: c89f45b831621cdc5e951694db44fecd8fab1150
Date: 2026-07-05T06:39:45.228Z
Layout: IDE
Build Type: Stable
Release Track: Default
Electron: 40.10.3
Chromium: 144.0.7559.236
Node.js: 24.15.0
V8: 14.4.258.32-electron.0
xterm.js: 6.1.0-beta.256
OS: Windows_NT x64 10.0.26200

For AI issues: which model did you use?

GPT-5.5-high

For AI issues: add Request ID with privacy disabled

41c14510-3380-4862-b4dd-b2efb343adde

Additional Information

This is consuming my Ultra subscription extremely fast - every basic question is suddenly a full context window - my subscription is currently burning at a rate of 20% PER DAY.

i.e. less than one week of usage - which previously lasted a month, with room to spare - is about to wipe out the entire sub.

Additionally: quality of Cursor’s responses is heavily degraded because it saturates the context window immediately and has to compact.

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

[Colin]

Hey there!

I’d be surprised if it’s sending your whole codebase. There are lots of things that might be taking up context, and you should be able to discover them by pulling up the Context Usage report in-chat after clicking on the context usage-ring.

For example, it’s very common for users to accidentally include rules/skills they aren’t expecting to, using up the context window.

You can even get as granular as specific turns in the coversation.

[A_A6 (A A)]

It’s almost entirely the conversation. I already checked that. I do not have any extra rules or skills.

As already noted: this is the SAME codebase the SAME agent the SAME configuration that - until last week - did NOT fill the entire context rapidly, but in the last 72 hours has aggressively over-filled the context.

[A_A6 (A A)]

Note: the only thing that makes Cursor useable at the moment is that CursorGrok is consuming very small amounts of context by comparison (it is doing what all other models did until last week).

[Colin]

Thanks Adam. We don’t have any other reports of GPT 5.5 consuming more context unexpectedly, either anecdotal or looking fleet-wide.

If you’re able to capture the Request ID of one such request with privacy mode disabled, I would be happy to look further!

Just a note that this isn’t the default GPT 5.5 model in Cursor, gpt-5.5-medium is. That can make a big difference when it comes to how much context the model collects / how “hard” it works on a task.

@A_A6

GPT-5.5-high