1. Codebase Interaction Tools:
• read_file(full_path: str): Reads the content of a specified file.
• write_file(full_path: str, content: str): Writes or updates content in a specified file.
• delete_file(target_file: str): Deletes a specified file from the codebase.
2. Code Modification Tools:
• edit_file(target_file: str, instructions: str): Edits a file based on provided instructions.
• parallel_apply(edit_plan: str, edit_files: list): Applies similar edits across multiple files simultaneously.
3. Code Search and Analysis Tools:
• codebase_search(query: str, target_directories: list): Finds code snippets relevant to a search query within the codebase.
• grep_search(pattern: str, target_directories: list): Performs a grep search using a specified pattern.
• file_search(filename: str): Searches for a specific file by name.
• web_search(query: str): Conducts a web search for external information.
4. Code Execution and Debugging Tools:
• run_command(command: str): Executes a command in the terminal.
• reapply(target_file: str): Reapplies the last edit to a specified file using a more advanced model if necessary.
These tools work together to provide a seamless and efficient coding environment within Cursor IDE.
Max does not use any credits. it uses USD. Maybe you posted in the wrong thread?
The things that confuse me about the recent changes re “MAX” are:
-
Seemingly odd price discrepancy
- claude 3.7 thinking costs $0.08 per call (now two requests vs one)
- claude 3.7 MAX costs $0.05 yet is the “premium” service option?
- how can the premium option cost less?
-
To think or not to think
- what happens if I have the thinking toggle “off” for:
- claude 3.7 max - how can it work w/o the thinking model?
- claude 3.7 sonnet - just bumps down to the “non thinking” model?
- o1 0 can it “think” or “not think?”
- what happens if I have the thinking toggle “off” for:
I use Cursor every day. My teammates and I love it. Just looking for some details on the above so we can gauge pricing / usage.
Just tried Sonnet 3.7 Max on one of my ongoing projects to review and let me know if a specific functionality was well architected. In the Chain of Thought, saw this and wanted to confirm…MAX can’t access/review codebase?
Running on usage-based pricing right now exclusively since I’m over the 500 Pro entitlement. Not sure if that makes a difference.
hi, first is every tool call a fast cursor request? And is there any cursor/rules content thad could reduce tool calling (like project folder, filesystem overview etc…)?
that seems incomplete isnt it? at least i see parameters missing and a few tools as well
@danperks can we have a “high” context mode for 3.5 as well? Perhaps not with all the other bells and whistles but essentially reading entire files/using 200k context windows. I understand the .05 per 3.7 query, but not the .05 per tool call. Is there actual overhead there? Is there a hybrid mode where 3.7 handles the main conversation and 3.5 handles tool responses or similar?
I’m frequently getting “The prompt is too long. Please shorten the prompt and try again.” using Max, even after 1-2 prompt inputs without much content.
@danperks have you considered making the pricing for Claude Sonnet Max more closely aligned with the actual costs incurred? If so, could you share why you decided on the current approach instead?
My understanding is that, under the current pricing structure, two otherwise identical projects—one in a single file versus another spread across multiple files—would end up with significantly different costs due to the extra tool calls needed for reading and writing to multiple files.
Wouldn’t it make more sense to align the pricing directly with the cost incurred behind the scenes?
@danperks @jake when I am using my own claude api key, I am still getting charged for the tool uses in 3.7 max. Is that normal?
What. a. disaster.
This launch should have been more thought out. Cursor, I love the innovation, but please respond to this growing thread with transparency, fairness and stability.
Do a survey, what are your thinking about the three billing models: token-based, query-based, and execution time-based?
Hi,
I’ve been using Cursor for a long time and always loved it, but after the updates starting from v0.45, I’m completely frustrated. I’ve tested this new version thoroughly, and honestly, I can’t say anything positive about it. My experience has been overwhelmingly negative.
Here are the main issues:
- Claude’s Performance : On simple tasks, Claude keeps “hallucinating” or giving wrong results all the time. This is super disappointing because it used to be much more reliable.
- Tool Errors : The tools keep throwing “Call time error” messages constantly, making them practically unusable. This is literally unacceptable. I get that the prompts you’ve set up for Claude might have some errors, but if the basic tools are failing like this, it feels like it goes against what Cursor promises as a product.
Cursor used to be a tool I loved, but since these updates, it’s become nothing but frustrating. I really hope the team fixes these issues soon because I’d love to go back to enjoying the app like before. But honestly, it feels like they don’t give a ■■■■ about what users are saying about these problems. Instead of fixing things, they just keep adding useless stuff to squeeze out more money, which doesn’t make any sense because people are going to stop using it…
I’m pretty sure at least 70% of users who have some coding knowledge (and aren’t complete beginners) have run into these same issues.
meanwhile… downloading Trae, paying Windsurf and bolt.new monthly subscrition, cancelling Cursor subscription… Honestly, I truly feel scamed, but seems not to be a refund policy.
I don’t really know why I send this message to the forum, if Cursor’s team won’t even read it lol.
I get your frustration, I am not affiliated with Cursor, but truly debugging an integrated LLM is almost impossible. The massive advancements in AI is going so fast that it is difficult to keep up. Everytime there is a new model, they have to go through a ton of iterations to get the model stable again.
Give it some time, they will iron out the 3.7 kinks.
I recommend sonnet-3.5 for now especially if you are a code viber.
If you have a well structured existing code base, 3.7 seems to perform better for experienced programmers that uses it as a assistant instead of code vibe.
Is this in cursor docs? Would be useful to see what tools the Cursor agent has access to.
Max mode Claude 3.7 prompts cost $0.05 per request + $0.05 per tool call
This does mean that, if left unchecked, it would could cost $10 if left to loop through all 200 tool calls it has available.
Am I getting this correct - even if we do not enable the usage-based pricing, we still getting charged if we (unintentionally) use the the 3.7 MAX model?
Correct me if am wrong, please!
