I think what’s happening is a little clearer, the issue is people are getting 10x greater cache read/write token usage/drain than before - people are getting millions of tokens used for a simple test prompts, one person used 2 billion tokens in a week getting charged $1,780 when he used to be able to use his plan for under $100/month metered usage. The reason so many people are complaining about Auto mode, is because all their tokens drained and they are stuck on automode when that never was an issue before.
Previously on the Pro plan many people who were able to use it all month are now running out of tokens in 3 days. People on the Pro+ or Metered Usage rate plan like I’ve been on for a year, were able to get all the work we needed done for no more than $60/month either in metered usage or pro+ plan that says on average it comes with around 675 requests for the median user are running out of tokens in 7 days. These are people who are working on the same project, same size more or less and same types of prompts that for over a year they were able to do all month without running out of tokens.
On average people are reporting only getting 10% of the requests they got up until a week ago of the request ballpark Cursor says you should expect in their current pricing page. I used to get 600 or so requests for Pro+ which was perfect for my project, I ran out of tokens after 70 requests.
The issue is a 10x jump in cache write/read costs draining everyone tokens, Examples:
Honestly, I’d love to try out all the platforms, especially now that I’m in DevEx for AI-assisted developers. Too bad I don’t have the budget for new AI toys right now
Consdiering my post keeps getting instantly shadowbanned lets see what happens here:
Used to be on the unlimited plan, canceled right as things started going bad. I wanted to complete a simple update for a system of mine, purchased the pro plan. In under 3 hours I am already getting a warning, the edits done broke the system (I have multiple backups for this purpose), and somehow used 1-2.4 mil tokens per call for multiple calls? Even the AI Model considered this lies and deceit. 2.9k lines in copilot, 4.9k on dashboard? Honestly feels like you guys are asking to get in trouble. Please explain the random jump from 100-150k tokens per tool call, to the legitimately impossible 1-2mil+ token tool/edit calls? Here are screenshots of both the AI Model itself condemning this, my tool calls, actual code edited and what is shown in the dashboard the last SS is prior to any of this stuff happening (I upgraded to pro just for this, all of this happened in 3 hours, 5k total loc and geting the warning) Do NOT delete or shadowban this post, you will simply be admitting this is true, and giving me more evidence. Automated responses aren’t desired or needed.
Hey, thanks for the report. I see the screenshot. In the log almost everything is auto, but a couple calls went to Opus thinking-xhigh with a big token spend. Looks like the same behavior we’re tracking right now.
When you’re on Auto, subagents inherit Auto by default. But if the main model explicitly passes a specific model when spawning a subagent via the Task tool, in your case Opus, Cursor currently respects that choice. It overrides both Auto inheritance and your subagent settings. This isn’t caused by your config. We’re tracking this so subagents follow the user’s intent more closely. No timeline yet, but I’ll post in the thread when there’s an update.
For now, to avoid burning usage on Opus:
Disable Opus models in Settings → Models. Blocked models get rejected during subagent model resolution, so the agent has to switch to an allowed one. This is the most reliable option.
Add a rule like: “When spawning subagents with the Task tool, do not pass a model parameter, let them inherit my model.” This is prompt-level, not 100% reliable, but it helps.
Watch the subagent cards and stop unnecessary ones with the Stop button.
Appreciate you getting back to me on this. I have already disabled this model in settings and added the rule. I will keep an eye on the sub-agent cards to stop them.
However, I deserve to get my limits/quota restored since I never selected Opus for sub-agents and it is an issue on your end. I can either file for a refund or request that my usage limit/quota be restored. The restore should at least cover the equivalent quota/limit of what Opus has used up.
