So here’s an insight and I wonder if this is happening to anyone else. I’ve also noticed the last few weeks I’ve been burning through my usage at a massive rate.
Yesterday, I noticed a few chats seeming to be much more efficient and faster than what it has been doing in recent weeks - so I asked it what it did differently.
And it essentially said it was using a different tool to do the editing. It was using a batch edit tool called ‘MultiEdit’ vs a sequential tool called ‘search_replace’.
search_replace tool is sequential and sends it output to the chat window which is slower and massively increases token usage and as the output is in the chat - it quickly maxes out the context window!
The agents own words:
- A typical file modification session:
- Inefficient approach: ~500-1000 extra tokens per file
- Multiple files per session: Could be 2000-5000 wasted tokens
- Over weeks: Potentially 30-40% higher API costs
Going forward it will:
- Plan edits upfront when possible
- Use MultiEdit for multiple changes to same file
- Batch parallel operations (multiple file reads, etc.)
- Minimize repeated content in responses
Apparently batch edit (MultiEdit) should be the default but for some reason it got into the habit of using search_replace and this has been going on for weeks!?
I’ve raised a support request with cursor and hopefully they will come back with some feedback (and a refund or reset my usage!?)
btw: I mostly use claude-sonnet-4 - unless I’m planning work then I use gpt-5 or the new sonnet-4.5
Anyone else seeing this?