Gemini 3 review after 1 month: inconsistent at best, poor at worst compared to Gemini 2.5

Google AI Studio
1

I’ve been using AI Studio and Gemini for almost a year, and throughout 2025, I was able to use Gemini 2.5 Pro without too many problems. The model performed well; there were a few drawbacks, but they were minor. The arrival of Gemini 3 was exciting because I believe a model can only improve by incrementally increasing its features, not by boosting some aspects while sacrificing others. Yet, that’s exactly what happened with Gemini 3.

My use case:

  • I primarily use AI Studio and Gemini for writing, YouTube scripts, articles, reports, but also for casual conversations like providing technical support for Linux, etc.
  • I don’t code and I rarely use Vibe Coding; the only times I use Gemini 3’s coding capabilities are for shell scripts.
  • Gemini 2.5 Pro wasn’t 100% accurate in its writing, but it was good at 90%.

And regarding Gemini 3 :

First of all, since Gemini 3 “thinks” differently, with a very strong emphasis on synthesis, I had to rewrite all my system instructions/meta prompts. I had spent months refining these meta prompts for Gemini 2.5 Pro, and I lost 10 days redoing them. On the surface, Gemini 3 does the job, but it’s disastrous because it’s inconsistent. One of Gemini 2.5 Pro’s major strengths is its handling of long contexts. Even in a conversation with 50 messages and replies, Gemini 2.5 Pro managed to keep up.

With Gemini 3, it’s not that it has the memory of a goldfish, but it’s very inconsistent. Sometimes it remembers, sometimes it doesn’t even understand what was said two messages ago. For my main task of writing, where the scripts are long and complex, it’s really mediocre. Then there’s the quality and creativity of the writing. Here again, it’s very inconsistent. Sometimes it works well, sometimes I feel like I’m seeing replies from ChatGPT December 2022!

The technical support has also declined in quality. With Gemini 2.5 Pro, I was able to solve complex problems on Linux, and one of the advantages of using AI for technical support is that you can get a truly customized solution. Sometimes I get mundane responses, but mostly, it veers into hallucinations that could damage my system if I don’t detect them in time. Many users, especially these days, use AI for technical support, and Gemini 3 is really bad.

So, in all my daily uses, Gemini 3 has declined in performance, long-term memory usage, and handling of complex problems. I get the impression that the model was boosted primarily for “vibe coding,” but not everyone codes. I’ve always liked AI Studio, the free version, and the customization options, but with such inconsistency, it’s not really worth integrating into a workflow. If you have to completely renew and modify the workflow for each new model, it’s a considerable waste of time.

3

Hi @Houssen_Moshinaly

Thank you for taking the time to share your feedback with us. We truly appreciate your feedback, as they help us continuously improve the AI Studio & Gemini experience.

Thanks!

5

Yep Gemini 3 Pro is terrible at coding. 2.5 was better.
It seems to be especially bad in the Agent context. It’s pretty much unusable.
It’s not just in shell scripting & similar it is bad.

6

It’s strange though, at launch it seemed to be extraordinarily, exponentially better than 2.5, though I have also noticed the same things reported here.

The worst part about it, to me, is the unbreakable stubborn behavior and the unwillingness to even use the search function unless you put the proof right in front of its face with the link. If I have to do the research and fetch the link to prove to the model that it’s disregarding important information, I’m actually working backwards.

7

I’d like to confirm these problems, hoping these are fixed on release or people might actually be in danger due to misconfigured prompts or misinformation. Overall. I’d rate it a 6/10 for now. It’s good at mathematics, mediocre at coding, but horrible at writing nuances in stories and context remembrance. There’s something seriously wrong with the model weights.

8

This was my experience also. It really was ‘oh wow, AI is finally here and GOOD’ for a few weeks. Around mid-December, it just totally became borderline unusable.

AI studio is slightly better, but it’s no longer a time saver and no longer viable. I’m definitely cancelling my Pro subscription. It’s a shame, but I imagine they had to sacrifice quality for cost and now it’s a real step back. They showed a glimpse of what’s possible and got the headlines and buzz to steal some of the thunder from OpenAI, but now they product they’re actually selling as been totally nerfed.

9

Yeah. I strongly believe Google somehow did lobotomy to gemini 3 pro just to make gemini flash look better.
Now cannot use anything except claude…
Gemini flash somehow nice for small changes compared to groks code model.
But I need pro for coding.

10

nice feedback, we need more of this for the google team to know the problems

11

I have had issues since the update to 3.0 pro. It got stuck, first in nano banana mode, and then switched to research mode and the content window locked, and do is research mode. All on its own. Result chat is inaccessible. I can only scroll 2 windows. No help from support. They didn’t give a lost more than a month of important work.

Now it’s memory is goldfish, can’t remember anything, don’t understand even basic tasks. Error 13, etc etc. Connection problems.

But when it works, it’s amazing. But thats only maybe 5% of the time.

Google you need to fix the issues. As a paying paying customer this is unacceptable. And get better support BC they are either instructed to be useless or goldfish them selves.

12

I have similar issues although I use Gemini for coding.

It’s hard to describe the experience with civilized language. I use Anti-gravity with Pro subscription and the fact that myself and many (if not most) of AG users are mostly using Anthropic models should be telling to Google.

Because of the rate limits, I use Claude only for analysis and for those times when I have to use Gemini for the same task, the difference is enormous. Especially Opus - it reasons logically while Gemini (doesn’t matter if 3.0 Flash or Pro) behaves like partially brain-dead. It takes inputs literally, does not consider intent. When something is not explicitly mentioned, it does whatever it wants, if it is, it still has problems with following instructions.

Also, it compulsively passes to coding even when it is required explicitly not to do so, and regardless if the mode is for planning or coding.

Basically, I use it only because I got tempted and bought the Pro subscription but as soon as it ends, I will move to another provider.

I don’t understand why with so much potential Google really lobotomizes their models :face_with_monocle: