Sonnet 3.7 thinking: We’ve made a few improvements to 3.7 prompting. Thinking will now cost 2 requests instead of 1. More improvements will be rolling out in the coming days.
Some clarification as to why this is an improvement would be appreciated. I’m currently still on 0.46.x so my premium requests are still at 1 credit per prompt and I’d need a pretty good explanation to convince me to do this update and start paying double.
How the ■■■■ do we turn thinking OFF? I tried 3.7 once, then tried 3.5 and it was still displaying thinking even in 3.5. The thought process is largely useless though if it’s hallucinating it might help to debug, IF there was anything I could do about it.
I think I killed cursor, restarted and at least thinking was gone from the default 3.5.
The improvements to Claude 3.7 have hopefully made it significantly more predictable and “tamed,” but Claude 3.7 thinking comes with significantly higher token usage and, therefore a higher per-request cost.
Rather than increasing the base plan price for everyone, we’ve chosen to adjust the request cost to keep the plan at $20/m while still enabling everyone to access Claude 3.7 thinking.
For rolling back, you can grab the previous version from Downloads | Cursor - The AI Code Editor, but this won’t reduce how many requests a Claude 3.7 thinking prompt uses.
@Cyber_ufo Regarding your screenshot, the stats are not instant, and your count may take time to update vs the request list at the bottom.
@hydrodog You should have them as separate options in the model list; if not, please make sure you are on the latest version.
My statistics are updated instantly, in my personal account I tried just 3.7 sonnet, 2 requests were removed instantly… and if 3.7 sonet-thinking = 3 requests =(
Are you telling me i’m paying double for software that’s worse. WTF. ■■■■■■ cheek of it. I did notice that disabling claude altogether, improved today. not by much though.
Insane that they are upping the usage cost of Claude when it barely works half the time. It genuinely feels like it has gotten dumber over the course of these past few weeks with little to none of the issues fixed that the community has been asking for since it debuted on Cursor. Very upsetting to see this. Between it not actually properly generating the code half the time and breaking when it applies it the other half, its really become an issue. Posts every single day about the glaring issues with these recent Cursor updates and the new models but the Cursor team refuses to acknowledge or look for a proper fix. Very upsetting and will probably be looking to cancel soon if this keeps up.
I took a yearly subscription and am certainly regretting it. Doubling the credits used is already pretty bad practice but insisting on calling this “an improvement” is really where it becomes too much for me. In addition to all other issues discussed here by other users. How many days after taking a yearly subscription can we ask for a cancellation and a refund? I’m UK-based. Since less is more you may consider sending me my money back “an improvement”
I’ve been having this feeling more and more, but typically it’s just because as the project size and complexity goes up, one shots go way down.
Just tried using 3.7 on an old project that o3 mini high fixed easily with copy/paste.
Claudes context window is small.. only so much complexity you can fit into it. I assume max performs better but I’ve not felt the value I use to feel.. again probably because my projects are getting bigger but ugh.. bummer. End up wasting more time trying to reconfigure rules and prompts..
I really do not understand you. If sonnet 3.7 thinking is more expensive, do not use it. As simple as that. Use the sonnet 3.7. It’s not cursor’s problem that the thinking model spits more tokens and thus is more expensive.
You like it. Use it. You do not like that model, do not. Disable it, remove it. Who forced you to use it?