Cursor is Burning Tokens Too Fast & Auto Model is Unusable

Ideas Feedback
, ,
1

I’m writing this as a serious user who genuinely wants Cursor to succeed. However, the current experience is becoming difficult to sustain in real-world development.

Main Problems

  1. Token Usage is Extremely Inefficient

Even with all optimizations enabled:

  • Include workspace files is ON

  • Context is carefully managed

  • Prompts are written as efficiently as possible

Despite this, tokens are still consumed too quickly.

A $60 plan is not enough to complete a single real-world project. This is not sustainable for professional use.

  1. Fixing Model Mistakes Costs More Than Writing Code

This is the most critical issue.

The models frequently:

  • Misunderstand requirements

  • Produce incorrect or incomplete code

  • Ignore constraints defined in prompts

As a result, more tokens are spent correcting mistakes than building actual features.

This defeats the purpose of using AI in development.

  1. Auto Model is Unreliable

The auto model selection:

  • Often chooses weaker models

  • Produces inconsistent results

  • Fails to maintain context properly

In many cases, the output quality is too low to be usable in serious projects.

  1. Context Handling is Inefficient

Even when:

  • Context is already provided

  • Files are included

  • Instructions are clear

The model still:

  • Misses important details

  • Produces redundant or conflicting outputs

  • Fails to follow structured instructions

This leads to unnecessary token consumption.

  1. Not Suitable for Large-Scale Projects

In real-world scenarios involving:

  • Multi-layer architectures

  • Backend and frontend integration

  • Complex business logic

Cursor struggles to:

  • Maintain consistency

  • Follow structured development patterns

  • Produce reliable outputs

It currently feels more suitable for small snippets rather than full systems.

Real Impact

Because of these issues:

  • Development slows down instead of speeding up

  • Costs increase significantly

  • Trust in AI-generated code decreases

  • Workflow becomes inefficient

What Needs to Improve

  • Better token usage optimization

  • More reliable and consistent model outputs

  • Significant improvements to Auto Model selection

  • Stronger context understanding and retention

  • Clearer visibility into token consumption

  • Option for a more deterministic coding mode

Final Thoughts

This is not written as a complaint, but as serious feedback.

I want to use Cursor as a primary development tool, but in its current state:

  • It is expensive

  • It is unreliable

  • It increases workload instead of reducing it

This is a real blocker for professional use.

If others are experiencing similar issues, it would be valuable to hear additional feedback.

2 Likes