In the last weeks, I have noticed Gemini 3.1 pro preview becoming SIGNIFICANTLY worse in the Google AI Studio Playground. It is doing errors that it was not doing before. To be more precise, what I have noticed is:
The model often fails to place the code inside the proper code blocks with the identifier, resulting in a file having 20% code inside some code blocks and the rest being fragmented as if it was in raw mode.
When this happens, I have to stop the reply, and I usually write something like:
“I had to stop your reply, because you were not placing the code inside markdown code blocks with dart identifier. Please provide again all the files, making sure that the code is placed FULLY inside markdown code blocks with the DART identifier, so that I can easily copy paste the updated files into vs code.”
And then the model proceeds to provide the file inside the proper code blocks.
Note, thas is also what I write in the original prompt, with those same words, but for some reason the first time it is often ignored by the model.
Another recent issue is that the model fails to number files it outputs properly next to file path/name
Eg it writes 1. for every file, instead of
1.
2.
3. etc
Another issue and maybe this is the most significant of all, is that the model does not always do the thinking, even with thinking set to HIGH, before outputting code. And most of the times when it doesn’t do the thinking part, it will make mistakes in the code output, which result in having to prompt the model again to fix those mistakes (endless waste of tokens and of usage limits).
Until some weeks ago, those issues never happened, this must be due to some recent change they applied to the platform or to the model. I would like google to seriously consider reverting those changes that have just made the experience and results of their platform/model much worse.