Opus 4.5 is great but pretty expensive if you're a heavy user. I still have one of the legacy $20 Cursor subscription with unlimited "auto" mode so I use that as much as possible, and when auto is struggling, I resort to Opus 4.5. Then, 50/50 chance Opus 4.5 one-shots it, or it also struggles just like auto. Then I have to think myself how to solve the problem (thank god I can still do that).
I'm curious after that subscription expires how much "auto" will end up costing me on demand...
I used to find auto to be bad but I'm more and more impressed by it. And when it hits a wall and I end up re-prompting dozens of time with useless output every time, at least it's free. Because this also happens with Opus 4.5 and I hate the feeling of paying so much for no usable output. When I try something with Opus 4.5 and it shits the bed, I fall back to auto quickly to limit the spend.
I think it's the reality of AI coding right now. You'll have the same issue with all the other tools. You pay for both good and bad output, and doubling down on bad output as many times as you want does not guarantee it'll make it good eventually. Yes, very similar to gambling.
i will try Claude Code with Opus4.5 on a similar MVP, if it still fails when session becomes higher that means i need to do manual orchestration. I wonder if there is already a pattern people use for that.
From my exprience also I agree that Auto on all the time is the way to go for now,
only pay extra for initial plan nothing else. Also avoid long sessions, after first context summarization Opus4.5 is pretty much done using Cursor, will check how others do and update here.
Tried both, openai models are also good, but something with agent getting out of plan after iterating for me, initial plan says to use xyz library and it uses, later on something goes wrong and fallbacks to a hack or simpler solution despite I correct again and again. I feel like summary of summary for the context causing the issue. Do you use it as snapshot task handling or MVPs? How do you manage long sessions?
Opus 4.5 is great but pretty expensive if you're a heavy user. I still have one of the legacy $20 Cursor subscription with unlimited "auto" mode so I use that as much as possible, and when auto is struggling, I resort to Opus 4.5. Then, 50/50 chance Opus 4.5 one-shots it, or it also struggles just like auto. Then I have to think myself how to solve the problem (thank god I can still do that).
I'm curious after that subscription expires how much "auto" will end up costing me on demand...
I used to find auto to be bad but I'm more and more impressed by it. And when it hits a wall and I end up re-prompting dozens of time with useless output every time, at least it's free. Because this also happens with Opus 4.5 and I hate the feeling of paying so much for no usable output. When I try something with Opus 4.5 and it shits the bed, I fall back to auto quickly to limit the spend.
I think it's the reality of AI coding right now. You'll have the same issue with all the other tools. You pay for both good and bad output, and doubling down on bad output as many times as you want does not guarantee it'll make it good eventually. Yes, very similar to gambling.
i will try Claude Code with Opus4.5 on a similar MVP, if it still fails when session becomes higher that means i need to do manual orchestration. I wonder if there is already a pattern people use for that.
From my exprience also I agree that Auto on all the time is the way to go for now, only pay extra for initial plan nothing else. Also avoid long sessions, after first context summarization Opus4.5 is pretty much done using Cursor, will check how others do and update here.
Yes! Cursor is still great for me. Try using Plan Mode or gpt-5.2-codex.
Tried both, openai models are also good, but something with agent getting out of plan after iterating for me, initial plan says to use xyz library and it uses, later on something goes wrong and fallbacks to a hack or simpler solution despite I correct again and again. I feel like summary of summary for the context causing the issue. Do you use it as snapshot task handling or MVPs? How do you manage long sessions?