Has there recently been a major change in how Cursor agents work or how their usage is calculated?
I started my new subscription period today, did some basic coding with the agent, and have already used 11% of my monthly allowance. I haven’t changed the way I use Cursor. If anything, I’ve been starting new chats more frequently than before.
Even a simple message such as “test” or “hi” shows around 100,000 tokens of usage. I disabled the high-quality models, set reasoning effort to low, disabled thinking, and tried Auto, Grok, and Composer models. However, the result is always similar: even a small task, such as editing three or four lines, can consume 300,000 tokens or more.
At this rate, my monthly allowance will not even last a week if I continue working as I did before.
Has something changed recently? Is anyone else experiencing this?
It reminds me of the usage changes Copilot introduced a while ago, which were one of the reasons I switched to Cursor. Now it feels as though I’m facing the same situation again. Is there perhaps a setting I’m missing that would restore the behavior I was used to?
Over the past two weeks, I paid three times the monthly fee in addition to the regular charges, but the data was consumed even faster and has now been suspended again. Eventually, I ended up coming here to Help. I have also realized that this is not just my problem. I want to request a refund, but I don’t know how. I don’t know of any way other than sending an email to the hi@ account. Anyway, I am certain that this is a problem on the Cursor side.
Hey @Yamaha_Fukoji, thanks for the detail. Nothing changed in how tokens are counted, but since ~August 24 Auto bills at the list price of whichever model it routes to (Grok, in your case), so it draws from your allowance. Details: Models & Pricing.
A short message like “hi” reads ~100k tokens because every message re-sends the whole chat’s context. Three things that actually help:
Start a new chat per task instead of continuing a long one - resets the context.
For routine edits, pick Composer 2.5 (not Fast) directly - cheapest per token.
Skip the Fast variants unless you need the speed; they’re ~2x.
Effort/thinking toggles only affect output tokens, so they won’t move the meter much. You can check the per-request breakdown on your dashboard under Usage.
@james7 - for anything billing or refund related, email [email protected] so the team can review your account privately.