Cursor Auto mode has become an inefficient, confused and borderline stupid model. It requires so much more context, redirection and reversion than I have had to deal with in the past. Feels like I am talking to ChatGPT.com instead of an advanced coding and development agent. It makes assumptions it has no reason to make, and often times almost ignores much of the context it is given.
Steps to Reproduce
Just by using auto mode for tasks that involve complexity.
Expected Behavior
Its behavior should be as it was a few weeks ago, where it can handle complex tasks and maybe needs some more direction compared to flagship. The difference between auto mode and new Claude models is like a toddler and an adult
Auto mode is designed to balance efficiency and cost. It doesn’t always route to the most capable model, especially for tasks it classifies as simpler. That can feel like a downgrade compared to flagship models.
A couple of things to try:
Premium mode: recently added alongside Auto in the model selector. It prioritizes quality over efficiency and draws from the API pool. If you’re on Ultra, you have plenty of API budget for this.
Manual model selection: for complex tasks, you can always pick a specific model (Claude 4.6 Sonnet, Claude 4.6 Opus, etc.) directly from the model picker.
If you want to stick with Auto but occasionally need higher quality, switching to Premium or a specific model just for the harder tasks is a solid workflow.
If you’re seeing Auto actively producing wrong or broken output, not just “less smart”, it’d help to have a specific example with a Request ID (chat context menu, top right, Copy Request ID) so we can look at what model was routed and whether something went wrong.
So is the price. I just use auto for boilerplate or for making similar code to what is already designed and implemented intelligently. It has its utility for how cheap it is, but can be unreliable for complex tasks.
With that being said, Auto is so cheap that I will often give it a complex task first just to see how it does. Usually its not that bad, especially if its not something super critical. Or do planning and initial build with Sonnet and then ironing out bugs with Auto or even gpt-5-mini.