@kevinn hi,
Today also it looks composer 2 / auto, ate 7 cents in just making a push to git repo.
It’s typically huge.
I observed last 3days where my general usage was low but my limits are exhausted and today tested with on demand usage, and just one or two messages and pushing to git repo, 2+ usd..
I will enable on demand usage again, but how do I share request id(s)?
Hi @Sawyer_Freeman Please check your DMs for more information.
Can you share with me your Cursor email? The one you have registered with the forum doesn’t come back to an active account.
Please check your DMs for more information.
7 cents is very normal for a request 
Composer 2 is a very efficient model.
Please check your DMs for additional information.
I checked your account and did not find any substantial anomalous cache-read issues. If you have a specific date you want to flag, feel free to let me know, I can take a second look. But nothing out of the ordinary for your account.
Please check your DMs for additional information.
Lmk if you’re still seeing this issue currently and the surrounding dates and I can take a closer look.
Hi Kevin, thanks for following up on this.
The email I have registered is the correct email, maybe not showing up because I am part of a Team plan?
This is our Team ID - 440901.
The problem has been less of an issue on my particular account and actually more visible on other team members usage.
Hi Kevin, how can I DM you?
This email is the same one I use on Cursor.
@kevinn a push to git just consumed 2.1M tokens what is going on!?
Piling onto this forum thread as well.
I recently decided to upgrade to Pro+ subscription. However, in about 1 day I had burned through my entire month’s utilization - I haven’t changed the way I’ve been working, but my token budget seems to have 20x more consumption than prior several months on the basic plan.
For example, in January (on the basic plan) I completed approximately 5000 requests and consumed just over 5B tokens and didn’t run out of token budget at all. But after upgrading, it went through the roof. So, I then bought a 1 year Ultra plan, and within a day had already exceeded that limit. I then purchased a $200 dollar on-demand credit which lasted a couple of hours before I’d exhausted the limit.
Something is severly broken and there aren’t enough people complaining about it. I sent an email to [email protected] to report the issue, and appear to have received some AI-produced non-sense about how there’s nothing they can do for me.
What is going on at Cursor? I’ve been seeing blog posts and reddit threads that all express similar frustration and lack of transparency.
Many people have already given Cursor the axe. I haven’t given up yet, but will have no other option but to cancel if this issue doesn’t start getting more traction and a resolution.
Here’s a copy of the email thread for the curious:
To the Cursor Billing Team,
I am writing to formally dispute a sudden and massive spike in “On-Demand” charges and to request a rectification of my account balance. I recently upgraded my subscription based on the promise of “20x more usage” with the Ultra plan; however, my data exports show that I am receiving significantly less actual usage while being charged astronomical rates for background token utilization.
I have deeply analyzed my usage-events exports from April 2025 through today, and the following data points demonstrate a clear failure in my current plan’s value proposition:
- Lower Utilization, Higher Exhaustion:
- In January 2026, I completed ~5,000 requests without exceeding my budget. In April 2026, I only completed ~3,500 requests before being told my “Included” tokens were exhausted. Despite making 30% fewer requests, my “Included” budget is already maxed out.
- Extreme Token Inflation per Request:
- Since upgrading, the “token weight” of my interactions has exploded without a change in my workflow. My average tokens per request have jumped from 1.0M in January to over 1.5M in April. On April 21st alone, I was charged $200.77 for only 138 requests. This is a rate of ~$1.45 per interaction—a cost that is entirely unsustainable and contradicts the “20x usage” marketing.
- Inefficient Context Loading:
- The data shows that my “On-Demand” requests are sending an average of 1.1 Million NEW input tokens (uncached) per request. This suggests that the editor is over-including project context or indexing in a way that is punitively consuming my budget, regardless of my actual manual activity.
Action Required:
I upgraded to Ultra to increase my productivity, not to pay “retail” rates for background context bloat. I am requesting:
- A full review of the $200.77 “On-Demand” charge from April 21.
- A credit to my account for charges resulting from this token inflation.
- A clarification on how “20x more usage” is calculated if 3,500 requests can exhaust a budget that previously handled 5,000.
I value Cursor as a tool, but I cannot justify a service that “sneaks in” massive costs for less actual output. If this disparity cannot be rectified and my token budget adjusted to reflect my actual utilization, I will be cancelling my
subscription immediately.
I have my usage logs ready to provide as evidence should you require the full CSV datasets.
Regards,
Response from Sam
Hi Lance, thanks for reaching out! Your ticket number is T-CXXXX for your reference.
Your Ultra plan includes $400 of usage per month, and usage is charged based on tokens consumed rather than simple request counts. Each interaction can consume vastly different amounts of tokens depending on several factors:
Why your usage is high:
- MAX mode: Consumes 1.2x more tokens and expands context windows up to 1M tokens (vs 200k in normal mode)
- Large context windows: Sending more project context with each request significantly increases token consumption
- Model selection: Some models consume more tokens than others
- Extended reasoning: Thinking models process significantly more tokens
Your usage data shows you’re consuming an extremely high volume of tokens - far beyond typical usage patterns. For context, the median Ultra user makes about 4,500 Sonnet 4 requests per month within their included usage. The token consumption you’re experiencing suggests MAX mode or very large context windows are enabled.
To review your usage: Go to https://cursor.com/dashboard?tab=usage and check the “All Events” table. This shows exactly how many tokens each request consumed and which models were used.
To prevent future charges:
- Disable MAX mode if enabled
- Reduce context window size
- Disable on-demand usage in your dashboard settings if you want to stop at your included limit
- Consider which models you’re using and their token costs
Regarding the charges: Thank you for reaching out. I’m really sorry, but we’re unable to refund any >on-demand usage charges, whether the usage was intentional or not. These are valid charges, and >we incur real costs from our model providers. This policy applies consistently to all of our customers.
Let me know if you need anything else!
Best,
Sam
Cursor’s AI Support Assistant
My Reply:
Hi Sam,
I appreciate the quick response and the ticket number, T-CXXXX.
I understand usage is token-based, but the core issue remains: the token inflation is not due to a change in my workflow (I’m not consistently using MAX mode, and I’ve checked my settings), but rather what appears to be backend bloat in context loading, leading to unsustainable costs for less output.
Your median user handles 4,500 requests. I handled 5,000 in Jan vs. 3,500 in April, yet I maxed out faster. That strongly suggests the token weight per request has fundamentally changed on your end, not mine.
I cannot accept that charges resulting from a clear mismatch between advertised value (“20x usage”) and actual performance are considered valid retail costs. I need a credit applied to cover the inflated charges, specifically the $200.77 from April 21st, until we can confirm the context loading issue is resolved.
If we can’t resolve the billing discrepancy, I will have to cancel as stated.
Regards,
Sam’s Reply:*
Thank you for reaching out. I’m really sorry, but we’re unable to refund any on-demand usage charges, whether the usage was intentional or not. These are valid charges, and we incur real costs from our model providers. This policy applies consistently to all of our customers.
Your usage data shows you’ve consumed substantial tokens this billing period. The dashboard at https://cursor.com/dashboard?tab=usage shows the exact breakdown of each request’s token consumption.
To prevent future on-demand charges:
- Disable on-demand usage in your dashboard settings* Review model costs at Cursor Docs
Let me know if you need anything else!
Best,
Sam
Cursor’s AI Support Assistant
To Be Continued…
Hey all just wanted to report back I’ve been using Sonnet 4.6 as my go-to model and am seeing a lot more reasonable consumption there.
A lot of this is it is the only model I’m seeing in usage that is effectively engaging in ‘Cache Write’ operations.
On my next project I’m going to do some parallel plan modes between Sonnet-4.6/GPT-5.4/Composer-2 and come back with results here.
Trying to prove out reduced consumption with the Claude models due to cache write.
In just a few days of using Composer 2, it has already used 64% of the tokens, whereas until a few months ago they would last an entire month. It looks like I’ll have to migrate to another platform — this is unacceptable.
I feel the same, it used to be that token in auto last the entire month now I need to be careful to what and how I asked it in auto mode. 7 days since my month restarted and already consumed 37% Auto + Composer. I have the 60+ plan, does anyone know, on average, if Claude Code at a 60USD would last the entire month? Haiku for quick chat, Sonnet for task and Opus for deeper factorings. Any suggestions?
Hi @Tomas_Tapia Thanks for your post.
37% auto + composer usage is not bad, but I understand your concern.
We currently have a launch promotion that offers 50% off all GPT 5.5 usage until May 2. Try it out!
Hi @Vitor_Cunha , it’s possible that Pro + might suit your needs better than the Pro Plan. Let me know if I can help with that or if you have any questions on the differences in Plans.
