Anyone actually noticed gemini 25 pro preview getting worse?

Has the Gemini 25 Pro Preview Deteriorated?

Recently, I’ve been experimenting with various AI models, and I’ve come to a surprising conclusion: my home setup is outperforming the Gemini 25 Pro Preview. Despite the advancements touted by Gemini, it often fails to meet basic expectations.

In my experience, the Gemini model has actually cost me around $200 in tokens over the past few weeks. This expenditure stems largely from avoidable missteps caused by its tendency to stray from straightforward instructions. For instance, it overlooks fundamental directives like:

“REVIEW THIS FOLDER AND SUBFOLDERS. You are looking for detailed information and examples for this project—they are in this folder.”

Instead of taking the time to consult these resources, the model forges ahead with misplaced confidence, ultimately returning with a flawed plan. Even more frustrating is its insistence that it has thoroughly reviewed the relevant documentation when, in reality, it barely skimmed through the initial prompts before veering off course and requiring additional refinement.

While I don’t expect a flawless performance, I do anticipate that the model should adhere to clear guidelines before concocting a solution.

I can’t shake the feeling that perhaps the current iteration of the Gemini model is being leveraged for purposes outside of coding, which could explain its disorganized knowledge base. The key-value (KV) cache is filled with irrelevant data that seems wholly unhelpful for coding tasks.

This line of thought took on additional significance with the recent model updates. It seems likely that the adjustments were necessary for the transition to a new build. I plan to revisit these issues in a few weeks, as I anticipate encountering similar KV cache problems.

In the interim, I remain hopeful that improvements will be made to enhance the model’s ability to follow straightforward instructions and deliver valuable insights.