I want to report a very frustrating experience with the Claude-4.5-Sonnet-Thinking model in Cursor.
On Apr 21, 06:50 PM , I made a very simple request: I specifically asked it to remove only this line from my home.css file:
margin-top: 64px;
That was the task. Nothing more.
Despite this being an extremely small and straightforward edit, the model used 498.5B tokens , which is completely unreasonable for such a trivial change. This kind of token consumption for deleting a single CSS line is not acceptable and makes the model feel inefficient and unreliable for real-world development use.
What makes this worse is that I was very explicit. I did not ask for a refactor, redesign, optimization, or any broader modification. I asked it to remove only that specific line, yet the token usage was absurdly high.
Please investigate this case. You should be able to verify it from the logs by checking:
- the timestamp: Apr 21, 06:50 PM
- the action performed on
home.css - the total token usage for that operation
I would really like the team to review why such a minimal edit triggered such an excessive token cost. This is exactly the kind of behavior that damages trust in the tool.