What you explain is forcing to make a Chain of Thought, it always “increases” models IQ.
I dont know if you use much .mdc and AGENT.md - that is basically necessary for complex projects. As far as keeping each file under ~600 lines.
i have been a life time supporter of cursor.. spent <$20/m for awhile.. then $100 then last month was $600.. i’ve switched to claude code..$100/m for max, and i have seemingly infinite opus 4.5, thinking mode… its not even close to compareable..
here was Eugene’s last response to odd billing displays in cursor AI on top of it.. I think Kiro has far clearer billing… constantly shows $ spent in bottom stat bar in vscode.
Thank you for reaching out.
I apologize for the confusion regarding Sam’s response. After reviewing the response, it appears to refer to your included usage, not On-Demand usage, when it mentioned the $203.37 figure.
Your On-Demand Usage billing is accurate. The total amount of your current paid On-Demand invoices for the current billing cycle is $280.71. With your On-Demand usage reaching $350, you still have a remainder of $69.29 that is yet to be billed as an invoice.
Your On-Demand usage matches your request history, which is accessible from the Usage tab of the Dashboard. I’ve manually checked, and you have more than 1.5k requests via On-Demand, which led to the incurred usage. In conclusion, your account is reflecting the correct costs based on your incurred On-Demand usage.
Requests are priced based on the tokens generated and the API rates of the selected AI model. You can view our Model pricing documentation for the full API rates breakdown: https://cursor.com/docs/models#model-pricing
I understand your perspective on this matter, but Cursor is using a usage-based pricing model. Your Pro plan subscription includes a guaranteed $20 in API agent usage and an additional bonus usage that Cursor provides at its discretion. Once you exceed this limit, the editor will prompt you to enable On-Demand usage or upgrade to a higher plan.
I also understand why you decided to switch, but I can’t really comment on other services’ pricing models.
Please feel free to reach out if you need further assistance!
Best,
Eugene
1 Like
kiro is fine, gives lots of credits to start (500 +50)… need to check out antigravity but idk.. cursor ide with free inline tab editing (i think its their OG goal) - with git commit summaries… is pretty solid with claude code opus 4.5 for $100/m.. it takes me 4 terminals to max out usage.. at any point in day.
Hey Nikita, Kiro is brilliant in its “Plan First” and “Hook Automation” approach. But it seems obvious to me that Amazon does not have a true commitment to the app. The update cycles are very long, and there is no active community. To me, it’s a bit like Amazon Video: it exists, and if people use it — fine. But putting hard work into it… no. 
2 Likes
I just downloaded Antigravity. I hit the Pro+ limit in 5 days, even though I was carefully selecting cheaper models to use.
I was tempted to upgrade to the $200 plan, but I saw many people saying they also hit the limit very quickly, which makes me think Cursor’s billing is “buggy” 
.
Maybe the number of tokens per request also increases with the upgrade.
I hope the upcoming Dynamic Context Discovery fixes that.
I have only hit the rate limit once in AG and that was during extremely heavy testing of my memory-journal mcp server. It took a few hours to lift the block. It is possible it was an auth-token problem and not even a rate limit. It is unclear. The ultra plan is on sale so is IMO unbeatable for now. AG also leverages Google’s search skills. It collects the conversations, notes, implementation plans, walkthroughs, task lists, takes notes (knowledge) and so on in the “Brain” folder and then uses this for cross-thread/project memory. This is quite useful. It also can edit multiple files in parallel far more quickly and reliably than Cursor. Regarding its MCP support, it’s not buggy as I previously thought. It is just using an older MCP protocol version. Problematic MCP servers typically call the wrong tool names seemingly, due to client differences, which is easily adjusted for with a rule or your prompt, etc. Also, the server instructions file is not auto-fed to the model but you can work around this by providing it in a rule or in a prompt, etc. I think all MCP servers could easily be made to work in AG with minimal adjustments.
I recently grew very frustrated with my developer and decided to fire him and just give the task to Cursor instead. I found Sonnet 4.5 needed a lot of handholding so I switched to Opus 4.5. It’s been amazing! Amazingly good, and amazingly Expensive! I tried several times to upgrade to Ultra but it seems my account doesn’t have access (can someone help there?) I started thinking of switching to Claude Code for that reason so I started to use cursor for planning and CC for execution. I don’t like the workflow but I like the cost savings 
I haven’t seem usage numbers like mine in any of the posts I read here so I feel I am overpaying … I’m sure I’m doing something wrong, besides being stuck on the $40 plan…
1 Like
I’ve hit my limit within 9 days.
1 Like
- Find a way to upgrade to Ultra
- How To Optimize Your Usage: The Best AI Models to Use, version 3.0
I just switched to the Ultra plan. Had been just paying for usage as I went but the anxiety of seeing that % creep up toward my on-demand limit was impacting my ability to get things done. I am working on my project part-time, in the evenings, after a full workday so I wanted to be able to maximize the effectiveness of my limited time.
After one day (about 2 hours) of usage (granted, mostly Opus 4.5 but also Sonnet), I have used 5% of my included API. That seems like it went really fast.
If one is using Opus/Sonnet exclusively (they are by far the best right now imo), is there any advantage to having Cursor Ultra vs the setup OP proposed (Cursor Pro + Claude Max used in Cursor)? I suppose to be able to use the Cursor agent chat, rather than working in the Cursor terminal, correct?
1 Like
Anti gravity is really only useful if you get it at a discount. Gemini 3 models are just really bad. I find them unusable in brownfield, a lot of mistakes are made.
Typically Opus is the really only usable option and the usage limits are severely restircted after just 2 days of heavy use they iwll have you wait for a 5-6 days.
The value you are getting from Anti Gravity is just not very good.
A better option would be Codex in that case if you would compare “Google vs OpenAI” because GPT 5.2 is significantly more capable
I have unlimited Opus 4.5 use with the AG plan. It’s fantastic.
Its the tokens that cause your plan to run out and bulge the costs. Straight tips:
- Reduce input tokens sent
- Reduce output tokens received
- Reduce thinking
However if your project demands high token utilisations, maybe nothing can save you. You’re probably better off with a $10,000 Mac studio 512GB M3 ultra variant, run GLM and minimax all you want.
Gpt 5.2, Gemini pro 3, sonnet 4.5 - is good enough for complex tasks
Literally exactly the same as me. I was getting up to $3,000 with cursors new parallel processing update. Switched to claude code yesterday. It is as good as cursor but just slower due to Claude Code not indexing the codebase upfront.