The problem of 「Gemini 2.5 Pro Preview 05-06 often not using thinking dialogs in conversations」 that I previously reflected in my post is now more and more frequent with the new version, especially when the number of tokens comes to around 10k, and it’s not at all clear how to get big models to think.
In particular, there are always temptations and incentives such as “I’ll reward you with a $200 tip” to faithfully complete a task or dialog.
I think Google’s current scaling back of model performance is getting to be intolerable ！！！！
3 Likes
In fact, after the update, even though Google saw the great outcry from the community, they didn’t add think buttons for Gemini 2.5 Pro Preview 05-06, their usual best-of-breed model, like flash.
This is actually extremely unfair to people who need research and learning intensely, because the unprocessed think answers are simply too bad to use, and there’s nothing to go on except changing their prompts endlessly.
That being said, Google is now excluding people from using the latest models for high intensity work on a large scale and large scale.
2 Likes
Hi @Cdkk_Wmfg ,
Thank you for sharing your thoughts. I’m passing this along to our internal team for their review.
@Cdkk_Wmfg Unlike the unhelpful response from the Google employee, here is something that will actually be useful: Google is adding a thinking enable or disable, and a thinking budget parameter, to 2.5 Pro in the API in June. It will work the same as 2.5 Flash.
3 Likes