So if I have it simply read 20 files, its 20 tool calls?
Why cant we use MAX for the prompt and then have it do tool calls under the cheaper model. Such a rip off. We already know the basic model can do tool calls… I just need max to determine what it will do… then have the cheap one do the tool calls. Is the auto model supposed to do that?
It’s crazy how Cursor is ruining its reputation day after day.
So you discovered that Claude 3.7 can manage the a bigger context better and decided to create another expensive MAX option rather than make the normal one better.
Even the files read are nerfed with the normal mode.
At this point I will consider to cancel my Enterprise Subscription.
This can only get worse.
Max will not help in helping the model plan out it’s changes, but only allows it to execute more across your codebase which keeping as much context as possible on hand.
You can use Claude 3.7 thinking to plan your actions, then disable thinking and get the model to do the necessary changes, all without ever needing Max mode!
Charging per tool use is truly a poor funding model. Reading a file in 200 line chunks instead of one giant read to cause 5 charges versus 1, or what not, is really user unfriendly. I understand the complexity of charging models, but you guys need another run at something where the costs relate to the actual costs. Not all tool uses are created equal, and the impact on LLM cost is highly irregular so should not be charged differently. You also dont want to encourage the limiting of using tools which is a great value to the final output. Charge per token like everyone else, or some other usage based metric + 20% or whatever you want, but keep it actual value creation not random “this seems ilke a good idea” models.
1 Like
Awesome. I woke up early this morning and told sonnet 3.7-max to implement a node-based polling service in my docker stack, listen to redis for commands, subscribe to external data feeds as a result, maintain an active list and post the data back to redis.
It got it 97% right in one prompt, while I was making coffee. And then we did some tweaking. It degraded after a long series of tweaks, I opened a new chat and it was fresh as a daisy.
Later in the day I downshifted to sonnet-3.7 regular, and it was like dealing with an intelligent person with brain damage - short term memory loss and ADHD.
I am so glad cursor introduced this tier – I understand why the cost. LLMs and cursor have evolved so fast that the pricing model that was built around GPT-3.5 does not cover the full capabilties of these newer models. I am glad we have a choice. If I am on the clock, or anyone working for me is, I think I would pay for them to use MAX because it saves most of those wandering soliloquies where non-max just responds to internal stimuli.
$10 spent on max saves at least $200 spent on developer time.