I’ve been using Gemini 2.5 Pro since around April, making it the 03-25 Preview, and I was initially very satisfied with the product.
Though the loss of visible thinking with the 05-06 Preview was an extreme disappointment (I chalked it up to the same “follow the leader” mindset that resulted in the loss of headphone jacks), I was still able to manage with some prompt engineering.
When 06-05 came around, I noticed two things.
- Its output reads as much more natural sounding.
- The first item is the only real positive.
I didn’t worry about this as it was only a preview version, and I figured that something better would take its place. 06-05 could manage desirable outputs with sufficient prompt tweaking.
Unfortunately, 06-05’s flaws carried over to the stable release without much noticeable change.
2.5 Pro stable feels grossly incomplete. It doesn’t follow instructions at all even with extensive prompt engineering, it’s grossly sycophantic (03-25 would actually push back if the user was incorrect whereas the stable version doesn’t at all), it’s frustratingly terse, and far less “lucid” than its preview predecessors. Every output feels like the SparkNotes of what I requested, as if it skimmed my input rather than actually reading it. 2.5 Pro has the tendency to flat out make things up, it refuses to think at high token values, and won’t even perform Google search operations after a handful of prompts even if I write concise_search in the prompt.
I bought the API thinking that I could use the previous versions of 2.5 Pro only to discover that they’re totally inaccessible, so I’m stuck with what feels like an unfinished version of a tool that I had come to rely on.