Feedback on O3 Behavior and Request to Restore Previous Mode

Feature Requests
1

Dear Cursor Team

First of all, I want to express my sincere appreciation for the amazing AI coding assistant you’ve built. Your IDE has significantly improved my workflow, and the intelligent support it provides is truly impressive. Thank you for such a valuable tool!

I’m writing regarding a recent change in the behavior of the O3 model. Currently, it appears that O3 is only available in MAX mode, accompanied by a very long context window being automatically sent with each request. While I understand the intent behind this change, it has unfortunately made the model much less usable for my needs.

Previously, your system smartly selected only the relevant context to send, giving me predictable token usage and clear control over the cost per request. That setup worked wonderfully for me—it was effective, efficient, and budget-friendly.

With the new setup, I sometimes end up paying over a dollar per call, without even knowing what exactly was included in the context window. This lack of transparency and control makes it impractical for me to continue using the model in its current form.

Please consider bringing back the previous O3 behavior as an optional mode, alongside MAX. Having both options would give users the flexibility to choose what best suits their workflow and budget.

Thank you again for your excellent product and for taking the time to consider this feedback.

Warm regards,
Yaakov

2

Second this. My first (and so far only) use of o3 under the new pricing model gobbled up $2.13 worth of credits for one prompt. It wasn’t a hard problem…it was a build error that any other model could have diagnosed. I only used o3 as an experiment to see how much it would cost.

3

tried one prompt with o3 as i had multiple tasks for the llm, o3 completed maybe 5% of what i told it and im being generous as it was a wrong solution for a rather simple php problem, ignored other instructions and tasks, consumed around 90K tokens, cost me 50 requests, tried to use python command that does not exist, let alone in a php repository, failed to use mcp server
none of these problems happen with other models and before 0.50 o3 was rather superior, i used it frequently for the exact same codebase

4

Bump on this for the Cursor team.

The request is to bring back a non-MAX version of o3 with limited context window, etc.