Gemini 2.5 Pro has consistently become worse in quality, I don’t know why these changes are being pushed to prod…Gemini 2.5 Pro 03-25 was the last best Gemini version…Even the April release was fine where the models thought enough and the outputs were great…05-06 model with adaptive thinking just doesn’t think when i need it to, its annoying because the outputs are actually lower in quality when it doesn’t think…The latest June release for Gemini 2.5 Pro June release although does the same amount of thinking as the original 03-25 but the outputs are bad
3 Likes
Welcome to the Forum,
Thank you for your feedback. We appreciate you taking the time to share your thoughts with us.
Could you please provide more details on the specific task where the 2.5 Pro (03-25) model outperformed the latest Gemini 2.5 Pro model?
1 Like
Similarly. I wrote an entire, enormous story with the previous Pro version—57 pages! It was as original as possible and preserved exactly the atmosphere I asked for. Today I decided to outline a different story and realized this model has become simply awful! It constantly describes things by negation and shows no creativity! I’m just disappointed in this release and would like to know if there’s a way to roll back, because this is just terrible.
Example from before the recent release (I needed lines for a creature that embodies the aspect of Honor in dialogues):
“I am the reflection of what was lost. I am the measure by which nothing is measured anymore. Call me Honor. I have come to behold those who tread among the ruins of the greatest of failures.”
“Father… Keeper of the Scales. He was given the greatest duty—to hold light and darkness in each other’s embrace, letting neither consume the other. But he fell. Not in battle, no. He fell under the weight of his own creation.”
“His children became the embodiment of extremes, and he allowed his family to become a reflection of the very chaos he had sworn to contain. He traded his duty for fatherly love, and this weakness shattered the very concept of Balance. His legacy is a cracked vessel. And you, Jedi, attempt to gather the shards without even knowing what you hold in your hands.”
Example of what the model produces now:
“Justice is not what is written in your Republic’s codes. Its purpose is to maintain the system’s equilibrium. Every action has an equal and opposite reaction; my task is not to pass judgment, but to assess how far your actions have shifted the system from its point of balance.”
“Your hesitation arises from seeking an optimal solution with no negative consequences. Such a solution does not exist. In this system there are no variables that lead to an outcome without losses. Your task is not to avoid harm, but to choose the outcome that brings the least evil and the greatest strategic benefit to the overall system.”
The replies completely severed the essence of the character I’d described to it; they’ve become utterly devoid of imagination and creativity.
4 Likes
The 1 million context window with the model output quality was very helpful, The new Gemini June release holds onto random context that is no longer relevant for eg… " This script is now tailored for your 0.0755% MAPE model from Phase 2B . If you later decide your SWA model (e.g., the one achieving 0.0655%) is indeed the one to submit, you would use the previous SWA-specific inference script I provided, ensuring the SWA model path and SWA loading logic are correct." , In actuality there is no model achieving 0.0755% in this current run, it was the initial context I provided to the model as logs and code when I started the chat along with clean problem statement etc, and I have progressed way past it since and the model was made aware of it and improved results through logs, etc… I had also pasted the logs and code for this chat too for it to realise that there was no 0.0755% MAPE in this run, so it is unnecessarily holding onto context that is no longer relevant…which is why I have to be extra cautious with the model to make sure I don’t chat with it, rather provide it with perfectly curated prompts so that I don’t mess up its context, with the older Gemini this was not a problem.
Also this new version is not as creative and has a lot more “You’re absolutely right” moments which I dislike…The older version was firm and sure about its outputs and instead corrected my approach/provided a reasonable explanation, if I questioned any part of its approach.
Even the 05-06 model even when it doesn’t think has better outputs when compared to the June release… The only issue with the adaptive thinking of the 05-06 model is that when it thinks, it clearly captures the bigger picture of the project, which is evident in the output, whereas when it doesn’t think, the non-thinking response is merely a continuation of the previous two to three chats and it clearly misses nuances, which makes it very unreliable, and trying to force thinking using prompts just works for 1-2 responses after which it doesn’t obey and just reverts back to non-thinking reponses.
5 Likes
It’s true, Gemini is even more sycophantic now, on a par with the grand dames that the OpenAI house has produced.
But we all know that we want a scholar who thinks deeply about dialog with himself, at least here at Gemini.
The latest update has made her lose her soul.
6 Likes
notice same issue too! very terrible answer for This GA model, I am disappointing for this.
4 Likes