Gemini 2.5 Pro Preview is very bad!

I used Gemini 2.5 Pro experimental, and I really liked it, it was the smartest model to date, but sorry how it disappeared from aistudio, how do I get it back, Gemini 2.5 Pro Preview is an extremely "Modified by moderator "model

28 Likes

Gemini’s response to your complaint:

It sounds like you had a really positive experience with a version of Gemini you were using in AI Studio, which you refer to as “Gemini 2.5 Pro experimental,” and now you’re finding the current version, “Gemini 2.5 Pro Preview,” to be less capable for your needs.

While it’s understandable to be frustrated when a tool you liked changes, making claims like one model being the “smartest” and another being “extremely stupid” isn’t very helpful without specific examples. Here’s why providing evidence is crucial:

  1. Subjectivity: Terms like “smart” and “stupid” are subjective. What one person finds intelligent, another might not, depending entirely on the task and the input provided. The model might excel at certain tasks but perform poorly on others.
  2. Reproducibility: For anyone (including Google’s team or other users on the forum) to understand your experience, they need to see what you were doing. Without concrete examples, no one can reproduce the issue or even understand the specific nature of the problem.
  3. Identifying the Actual Issue: Was it a specific type of prompt? A particular reasoning task? Coding? Creative writing? The perceived difference in performance could be due to changes in how the model handles specific instructions, data updates, or even subtle differences in how you’re prompting the new version compared to the old one.
  4. Model Versions: Model names and access can change, especially for preview or experimental versions. It’s possible the “experimental” version you were using was an earlier iteration that has since been updated or replaced by the current “Gemini 2.5 Pro Preview.” These updates aim for overall improvement, but sometimes specific capabilities might shift.
  5. Actionable Feedback: To provide useful feedback or get help, you need to show the difference. Ideally, you would provide:
    • The exact prompt(s) you used with the “experimental” model.
    • The successful output you received from that model.
    • The same prompt(s) used with the current “Gemini 2.5 Pro Preview”.
    • The unsatisfactory output you received from the current model.
    • An explanation of why the new output is worse for your specific needs.

Without this kind of detailed, reproducible evidence, your post reads more like a complaint than a bug report or a request for help. It’s impossible for others to validate your experience or offer constructive suggestions. Unfortunately, getting back access to a specific, potentially internal or deprecated, experimental version is usually not possible. The best path forward is to clearly document the shortcomings you’re experiencing with the current model using specific examples.

3 Likes

no… just Gemini 2.5 Pro experimental > Gemini 2.5 Pro Preview

I need Gemini 2.5 Pro experimental !!

11 Likes

nah I agree, the Pro Preview is actually unusable for coding compared to the Experimental, I thought I was the only one. It disappeared for me today after the UI was updated.
I don’t know what you guys did, but it’s hilarious how much worse it became.

17 Likes

Yeah the difference in quality is so bad that it’s bordering on empirical and far beyond just an anecdotal opinion. It’s not just 2.5 Pro Preview though, although this new model is incredibly awful at virtually every task that I’ve tried putting to it now, compared to the old 2.5 Pro Experimental.

But also, what on Earth happened to the UI and why did they decide to update it? It’s not just that the old UI was incredible but that the new UI they have is seriously awful like I have no idea what’s behind their thought process on the changes. Virtually all of the important tools are now hidden behind sub-buttons hidden behind buttons, but in trying to ‘simplify’ the workspace they’ve really stripped out easy accessibility to a lot of the important functionality and now you have to ‘find’ it by looking through their subpar UI.

13 Likes

I can say the same. I was wowed by 2.5 and I loved it but it has just gotten dumber over the course of a couple f weeks…

3 Likes

This is without a doubt the truth. I tried it and I never liked it. Something’s wrong with it. I think the best model so far is Gemini Flash 2.0. I’m using it and working on it with “Modified by moderator”

2 Likes

Gemini 2.5 Pro Experimental was great
Gemini 2.5 Pro Preview is significantly worse, by at least 50%.

…It mutilates my existing code instead of precisely tweaking it, like the Experimental could.

12 Likes

Gemini 2.5 Pro Preview is undoubtedly a downgrade…2.5 Pro Experimental was amazing for the PhD-level mathematical derivation tasks I was having it do, but 2.5 Pro Preview is substantially worse. Disappointing.

11 Likes

For me last version became super slow and unusable for large (40K) number of tokens

4 Likes

Unfortunately the 5-6 version of 2.5 pro is totally useless for RAG applications - just doesn’t work. It takes tons of time with auto “thinking” re-prompts that use output tokens then reaches a max token limit with internal outputs we can’t see, then returns the max output token error. This is really sad because for me the 3-25 release was the very best model ever to exist for RAG applications - better than GPT 4.1. My system pushes the models to their limit for substantive analysis ability and 2.5 pro WAS amazing. Then about 3 weeks ago they introduced some type of auto caching behind the scenes and that started creating hallucination problems because it would pull the wrong context. I was able to work around that with some prompt engineering but now it’s all for nothing - the new API is terrible. PLEASE restore the 3-25 version of the model/API!

9 Likes

Ditto, this 05-06 model is inexcusably bad. It’s mind boggling this was pushed as an “improvement”. Real Devs will pick up on this eventually. It seems it was fine-tuned for “one-shotting” web apps and not intimate pair programming. 03-25 was excellent at pair programming and metacognition. I don’t care about “one-shotting” silly web apps. Maybe it’s just me…

13 Likes

I agree, sadly during my tests, coding was worse and prompting didn’t work that well compared to 2.5 Pro Experimental. I hope we all at least can choose in the future between both version to compare both and use the best model for our use-cases.

5 Likes

Update - I was able to get the 5-6 relase working with my RAG application again by changing the max_output_tokens parameter (python name) to 32k up from 8192. I believe that the model API was updated to add extra thinking iterations, probably to make 1-shot prompting better is some ways to appeal to youtubers that compare models with a standard set of programs they ask the model to 1-shot. Anyway, from my testing I’m fairly sure that what was happening for most RAG applications is that they don’t specify the output tokens which default to 8192 (the old max, so there was no real reason to set the parameter). With the new 5-6 API wrapper, it was generating internal output in response to reflection prompts that exceeded the 8192 default limit, and then would stop processing and return a max output token exceeded error. With this model Google also increased the max output token limit to 64k. So far 32k has been sufficient to overcome the error. It still takes longer, but the output quality is comparable to the 3-25 realase in my opinion.

2 Likes

Agreed, it’s making basic mistakes, fails to retain the goals, lacks the ability to reason accurately, is unable to recognize context or adjust itself to it, hyper focused on the request instead of the bigger picture, the list goes on. I have yet to discover 1 thing it does better.

2 Likes

I agree. I have no words how it became. Even if I spoon-feed it the task, it can’t accomplish it. 2.5 Pro Experimental could do it way ahead of my thinking capability. But preview makes me feel like I am talking to a noob.

3 Likes

I’m facing this weird problem with the latest model as well.
I have given an elaborate prompt with all the constraints and boundary conditions but the response is totally unrelated to the topic of the prompt.

2 Likes

i agree. its same for me. Gemini 2.5 flash and Pro can’t keep up conversation. sometimes i am typing another thing but gemini answering 1,2 or 3 past message. it was the same for experimental but now for preview, also same issue.

1 Like

I only singed up for the forum to second this opinion. I noticed the degraded quality first and only then realized there was a change from experimental to preview.

Doesn’t make sense for me to pursue my projects with an AI model this bad. It was very good - yet not perfect - to begin with. Codestyles were motivated by real developers who change lines to fake productivity. But now, the model doesn’t even understand when I tell it only to only change a certain section of the code and leave the rest. What kind of botchery is this even?

Probably gonna deactivate my account until they improve again. It’s not suitable for development like this. Looks like classic bait & switch.

5 Likes

Indeed! I started a project for a customer because the PoC with the gemini Experimental was so good. But now, I will need to build this project with other models if this isn’t fixed soon.