I see the announcement says, costing $0.05 per prompt AND $0.05 per tool call!
So, with one prompt (request) and letting the model iteratively fix linting errors, we’ll likely go bankrupt, right?
Oh, I see another linter error → $0.05
Oh, I left another type issue unsolved → $0.05
Oh, I forgot that other thing → $0.05
I just signed up for pro and they immediately rolled out a new model that isn’t available for pro users. Absolutely shocking.
I said a few things to the AI prompt that it declined to convey to the marketing team. Feel free to use your imagination.
I’m satisfied with 3.7 MAX but the linter errors part is really a weird one.
Linter errors are produced by the model not the users but each time linter error occurs, the users are charged. My dashboard shows that most of my 3.7 MAX cost comes from linter error tool calling.
Simple answer. Yes.
Max is a success switch. Wanna make your project a hit? Flip it on!
I don’t think any official people will answer this question. 'Cause they always block your posts. Soon, no one will question what Cursor does 'cause they won’t be able to see those questioning voices. “bury one’s head in the sand”, which is exactly about this kind of thing.
The cost is ridiculous compared to 0.46, especially we cannot even choose to stick with 3.7 thinking instead of max mode right now.
Edit: I was wrong about the thinking mode as pointed out by @AbleArcher. And after check, I found we could still stick to 3.7-thinking with tool use with expense as “just” twice as non-thinking 3.7 if I understand correctly. It is not ridiculous but as heavy agent user, I will cost much more than working with cursor 0.46.
request cost is reasonable. request per $0.05. but i do not agree with ‘premium-tool-call’ cost. it charge you $0.05 also.
if one request call premium tool 10 times, then it charge you $0.55 per just one request.
and i think, it does not have $0.55 value.
Yeah, you’re probably better off with RooCode / Cline with direct api calls. I’m not even gonna try , funny they even declare it as MAX without’s anthropic permission.
It’s literally just non-nerfed context length sonnet 3.7 with thinking
The cost is ridiculous compared to 0.46, especially we cannot even choose to stick with 3.7 thinking instead of max mode right now.
The model selection has changed.
claude-3.7-sonnet with Thinking enabled should be identical.
If you turn on thinking, all non-reasoning models are greyed out at the bottom, selecting one of them disables thinking.
As mentioned in the forum thread, Claude Max is likely not for the average user here!
This is really intended for people who:
a) Are not concerned about the cost, and are rather willing to spend $0.30 but to increase the odds of Cursor being able to complete a request in one go
b) Have a situation that requires the last 10% of intelligence that is unlocked using these enhanced parameters
c) Who do not want to spend time setting up Cline, or generating API keys, and loading balance on their APIs
It’s a one click switch on purpose, so that people who just want something done can have the best odds of doing so, without having to set anything up for it.
We still think our base model implantations are still the most efficient solution for using these models without massive cost!
Regarding the tool call cost, we’ve chosen this pricing as it’s the closest concept we found that correlates to the actual price of these requests.
We wanted to make sure the cost of each request is transparent, so you know how the costs are transparent!
@danperks
Even though we should be careful on our end, it would be very good if Cursor could ensure with utmost care that billing is happening only on serious API usage.
For example, my very first try with Claude Max just now led to this kind of behavior:
and this:
This was a ~$1 interaction, but it delivered no more value than regular Claude. I know that you’ve only just started offering this new service, but I think you should be super careful about not overbilling!
26 premium tool calls * 5 cents per tool call ONLY IN 1 PROMPT CHAT ?
SO expensive! but not helpful
I think this should come with some warning in the prompt, or at least some estimate on the price it will charge. or maybe a cap to limit spending.
We don’t even know what to expect or if it’ll perform better or worse than other models.
Perhaps the conversation model should propose if using MAX is likely to give a better result?
No one knows how much you’ll pay until the chain finishes!
Yes, they disabled Github Copilot Chat in version 0.47 to introduce these higher-priced features.
just used roo cline- way better than using cursor- it was like 2.50 a page, (5 components/functions) but actually quite nice.
You guys understand that the model can iterate infinitely on linter errors, and sometimes it introduces errors in the code that it will iterate again to fix, right?
yep It did happen to me a few mins ago, tried to do a quickfix to test the max, it did a mistake and spent 95 toolcalls to fix its own mistake, which it didn’t at the end, so be careful
about 100 or 150 for me last night. Canceled my sub today. going to try another ide.
what ide？？？
After spending over $50 in just a few hours, I am greatly disappointed with Max Sonnet overall. The issues I encountered are exactly the same to those with the regular 3.7 Sonnet and Thinking in the Cursor Composer. It misses obvious information, does not adhere well to specified rules and in-chat instructions and must be reminded of them constantly, its efficiency is, in my experience, very similar to regular 3.7 - the agent needs to be corrected all the time to not deviate even from the accepted implementation plan, that’s already pinned in the conversation context. It has a strong tendency to “simplify” my implementation, which usually (almost always) ends up being a code loss that I simply rejected at review, while my problem still isn’t solved, so it’s a pure net waste of time. I wanted to give the new feature a fair shot, but I will not be using the Max option again.
Also, it was really annoying that even though the developers claimed that the Max option unlocks the full 200K token context window, the chat was constantly prompting me to open a new conversation to maintain high performance, in intervals, I would say, that are exactly the same as when using the regular 3.7.
In summary, I absolutely do not see the Max option to offer me a Smarter Sonnet, only more eager to use tool calls which are charged separately. The effect was greatly disappointing. I am convinced that if I were using a regular Sonnet in the Agent mode, I would basically get the same result without being charged additionally. The option will absolutely bankrupt you, giving very little in return.
To clarify, I would be absolutely okay with the cost that comes with the Max option if it would offer equal value in return. Currently, it absolutely does not.
