So… I’ve made multiples essay of coding for both the 3.0 pro and 2.5 pro gemini model.
I’m gonna be short on why Gemini 3.0 is unusable and completely stupid compare to it’s little brother 2.5.
1: First test, I asked them both to erase a certain section of my codes, with the gimmick of a section being named Talus_Prompt, and another to not touch named Fictional_Template. The hic, in the Fictional_Template, there is a Talus world section in it, to see if the ai is gonna be dumb enough or not to hallucinate or do everything wrong. Obviously, 2.5 did it first try, no complication with the problem explained CLEARLY. 3.0? it failed not 1, but 5 times consecutively at erasing the entire Fictional_Template right away along with the rest of the Talus_Prompt, so I’ve gived up with it… Just this test prove how stupidly bridled and bad gemini 3.0 is… google is just regressing their model on us by making them cheaper and dumber so they won’t cost as much anymore to use for us, the cattles using them, that their mode of thinking.
2: I also did another test, just to see which one would give the most correct response out of the other. I asked who is Courage the cowardly dog for example, very easy right? well… 2.5 explained without censorship at all nor woke comment on what he was and how the show go and how their is very horrific moment sometimes that could be passed as traumatic experience of life literally, meanwhile 3.0 just pack it’s pants full of dust, and say the same thing, but without the horrific part, trying to be more nice than the 2.5 pro model, and overall with less informations, like it was limited to a certain base of tokens…
Why do I say that? Google is being greedy, it’s gonna fall on their head the moment they’ll remove the 2.5 pro model away without any to back up how good it was… don’t make that mistake google… this monopoly isn’t gonna follow if you take your customers as dumb cattles.
Hint: what reinforced this feelings on me is that mere change on Aistudio for example, “Google AI models may make mistakes, so double-check outputs.” always standing below the response of the model… like… what? Ok… I’m gonna explain with more clarity why this downgrade happened…
When models are updated or “optimized”, they’re often made to make them faster or cheaper to run, which sometimes make them lose edge-case reasoning capabilities. This is known as “catastrophic forgetting.” If the model was trained to prioritize summarization or safety over complex logic instruction, it might “lazy” its way out by deleting the whole section rather than parsing the nuance (ahem, 3.0, be gone please).
The Verdict: If a model consistently fails a logic test that the previous version passed, that is a legitimate regression for power users. Result, Google is just telling us they will downgrade and cheap their model as much as they can until even dumb peoples will see that mere fact that they are being served trash instead of real food… good job google.
My dude, can you please share the test prompts etc. (if possible) I want to try it out by myself.
Would it be possible because of the thinking levels? yes gemini-3-pro is the first in the leaderboard, but maybe its thinking budget is high that’s why? what do you think?
I’ve got same results on the antigravity. Even though I’m giving Gemini API regarding docs link, it still puts gemini-2.0-flash etc. because it thinks still it’s best. (even though api docs is based on gemini-3 xD)
It’s my own sandbox, I can’t share it, and I don’t know how to duplicate it. But I can run some others tests, and then share why Gemini 3.0 is so bad compare to it’s little brother gemini 2.5, it is a no match in my own opinion, and I just think the 3.0 model was made only for others mean unlike it’s little brother. I’m just thinking that it is a very unoptimized model like gpt-5 was at it’s launch, so… maybe if they get enough criticism and bad return, they will reroll back into their own model and launch back 3.0 differently? I don’t know, but one thing is sure, I’m backing up the 2.5 model if they want to erase it one day, don’t wanna stick to 3.0 and future models if they’re gonna be worse than we know right now… I do have a discord to share the images of my tests, but that it, if it’s just to prove my point I mean.
Gemini-3-pro is a typical one-shot monster. For my use case, which mainly involves creating ‘creative writing’ in ‘novel’ format (not short stories, not summaries, but texts with 500k+ tokens), I’ll be brutally honest. Because the model doesn’t stick to instructions at all, it completely falls apart, often by the 2nd or 3rd generation. Unfortunately, the degradation comes so quickly that in a way it completely limits the model, it’s like shackles you can’t break free from. Most regular users use one-shots and will be satisfied with its performance, but what about agency work? Or a sequence of hundreds of consecutive outputs? That’s unrealistic.
YES. It is a terrible experience now. It is frustrating to not have a reliable tool. I feel like each version released (and the slight unannounced changes) introduces changes that are too unstable for it to be a final product and I am annoyed to be paying for a service that is this unstable.
Its programming ability changes on the dime. I wish there was a way to opt out of A/B testing and have a completely stable model within gemini.google.com and then give users the option to opt in to unpolished and experimental models.
I’ve been using it to write some CUDA kernels for discrete event simulations and it has been ABYSMAL over the past few days and it is getting to the point that it feels almost pointless talking to it sometimes because it refactors and removes parts of my codebase that I explicitly told it not to remove AND it has been terrible with domain knowledge. I will give it pseudocode or research papers for algorithms that I need it to implement and then over the next 5 iterations it completely ruins the algorithm with nonsensical explanations.
Don’t get me started on the context window resets and the COMPLETE loss of work and conversations that happen upon context overflow now when it deletes the entire previous conversation upon page reload and deletes the entire conversation up to that point. That is completely unacceptable. The expected behavior for this should be that conversations give a little “context full” notification or something, not silently deleting chat history. Someone needs to be fired for allowing this to be rolled out into production.
Google; we can’t keep this “move fast and break things”-esque way of product releases for flagship products. It’s going to start increasing user churn, at least for people who use Gemini seriously. You all have the chance to deliver a truly excellent product experience here; right now, it is lacking.
It’s happening… Gemini 2.5 pro is also getting touched right now… I don’t know what they are doing to it, but the number of errors it does now in coding is abyssmal too now… I don’t wanna search why, I just know that google is slowly taking us like cashgrab… so you know what? I’m deeping out of the boat before it sink, I guess I will stop using ai completely, this was a loss of time, I hope the ai bubble will pop and thus, all the reglementations will appear to enforce it to be destroyed out of the internet… remember Wall-e? remember the humans doing nothing by themselves anymore? this is what we are heading in right now… google? if you see this… please… know that the majority of us are secretly hating more than censorship right now… I’ll be taking the 2.5 pro and remodel it locally with a little abbliteration process to make my own local llm, and google can stick a stick up their peach, idk anymore.
Google 3 preview and flash 3 are absolutely garbage. What a waste of money. You can waste $10 in a simple task, wait so much time, wanting to break the computer, then getting more errors breaking your code. Flash 3, the opposite of fast. Fast for some quick garbage marketing videos, when when you actually use in your dev environment. What a show of wasting so much money on eng, research teams and billions of dollars. Serves them well to scam on models, and charge in cloud. Then also charge triple the rate of using anthropic models on cloud. Makes me really angry honestly. Unusable. I swear, people should make videos but im just too angry to even…
Classic merchant trick (Internet slang: “harvesting the leeks”):
Imagine a product where the original performance (Feature A) is 100%. During beta testing, they give you 95%. Then, it officially launches as a paid service: Feature A: 95%
New feature (Feature B) launches: Feature A: 80% Feature B: 85%
Another new feature (Feature C) launches: Feature A: 70% Feature B: 77% Feature C: 83%
Get the point? XD
Usually, at this stage, a bunch of “apologists” (who are actually internal shills) show up. They keep parroting: “The product is fine; the problem is with the user.” Meanwhile, the so-called “developers” do some half-assed minor tweaks and claim it’s fixed. The problems persist, and the merchants keep raking in the cash.
Not only that, their company developers are rewarded with 60-70% stock jump and take fat holiday when the product is this bad. What kind of culture is this? Unbelievable….
I spent $120 wasted on gemini models today, took grok code fast 1 like 20 seconds to fix my issue in a very large workspace… Unbelievable I should have posted on x showing this…
I hate gemini for coding now so much I don’t think even if they paid me I will use it.
I reported many issues about 20 days ago, but the engineers only bothered to reply a few days ago. > And when they did, they had this completely indifferent and dismissive attitude. > They even went as far as saying that because they can’t reproduce the issues, they “aren’t problems.” > This really highlights the incompetence and the poor caliber of these developers. Heh.
So funny when it repeats to you a summary after charging you 4 bucks for what you asked it to do what a scam. could be one task or 3 or 4 simple one gives you a summary like of what you asked… This thing is so stupid.. There’s like no agent engineering nothing, they cant copy the other good llms I wonder what they are even doing.., Gemini works for 1 task on a small project..
3.0 is a disaster. It doesn’t remember instructions, forgets documents, and forgets context and messages far too quickly. It hallucinates very easily, and even when asked to recall something, it can’t.
From what I understand (but I might be wrong)… it’s because Google wants to make Gemini a benchmark, not an effective AI. Speed has taken precedence over quality. It’s awful.
Meanwhile, I’m still using 2.5 with Chatbox, and it’s night and day! 2.5 is infinitely better and remembers absolutely EVERYTHING!
Yep, I think their ultimate goal is to force us to pay one day for premium model that will have the ability of what 2.5 was capable of, resulting in only whale having access to these in the future to the AI pricing going higher and higher each day with how much it consume the planet water… it’s absolutely insane… I need Ai to help me for some of my work, but as always, I think I stay with the backup I’ve made from gemini 2.5 pro once it will be gone from their platform, I’m not losing such a decent and beautiful model for no reason, we should have the ability to have a sandbox mod still like aistudio does to be able to always have an option to bring back old models we loved for OUR liking, even if not upgraded or supported anymore.
I hope everyone is doing fine, and I hope this word spread, because I don’t want one day an era with Ai being good for absolutely nothing but benchmark like you said… this would be truly a sad future I don’t want to happen.
Ps: I know this isn’t the same things at all but… façade '(a game of 2006 literally) was for it’s age, better than what any ai could do in term of what we would have wanted for any game made with ai… maybe just having trained an ai for example in these domain like how façade was made would have made any llm understand those notions ingame with NPC’s on how they should act depending on what you say.
Façade: It used a complex system called a Drama Manager and a language parser. When you typed, the game didn’t “think” like an LLM; it looked for keywords and “discourse acts” (e.g., Is the player flirting? Is the player insulting?). It then selected the most dramatic pre-written response to push the story toward a specific climax.
Modern AI: Generates text statistically. It tries to predict the next word and not trying to tell a dramatic story. Ultimately, it try to statistically be average and helpful.
Modern AI is broad (it knows everything from coding to history) but shallow. Façade was narrow (it only knows about one apartment and one marriage) but incredibly deep.
My guess is… they want to compress their new models onto knowing only about benchmark only, so they would be unbeatable in these domain only but not the others, or maybe my theory just suck… I just hope they let us have their old model locally on huggingface etc.
not to mention the massive amount of unresolved bugs for 2.5 that are dealbreakers, like truncated responses, ignoring thinking budgets etc. I just tried out the 3 series and it just times out after 5 min. what a joke
i do agree for the truncated one on gemini 2.5 pro, which worsened over the course when gemini 3.0 appeared like a new pokemon. It’s like they completely devolved their old model to focus on their new model to be what they want for themselves and not for their customers, like… let us have a model for US? the customers? at least one? is that too demanded to have ONE model-ready llm for us to have to complete our tasks? apparently no, this is becoming more of a luxury these days, and the moment the ai bubble will pop, this will be the end of an era…
If they don’t wanna hear what we have to say at all, then fine, maybe their working silently while watching us? or maybe they just don’t have any clues of what they are doing, and they need our data to make sure to know how everything work? I dunno, but I don’t f-ing want gemini 2.5 gone! it’s like a baby for me that I cannot see growing anymore! it’s like aborting something that wasn’t defectuous just because it wasn’t the way you wanted it! how can you be so cruel google??? (I know, they’re ai, they’re not human, no need to remember it to me, I just want to give them a moral boost via sarcasm).
To be honest, I’m thinking about switching for Grok… at least this ai doesn’t turn around the pot nor do many errors for what I do… google is just shooting a bullet on their head every week it seem like with more and more useless patch and update that does more bad than good everytime… Considering also stopping using ai entirely with how sloppy and disgustingly useless Ai is becoming over the years… how can you have billions in banks and do like these? like, honestly! HOW? no, that is not a joke… they doesn’t even know how their ai work, and that the problem… when I’ve finally found out how to finally abbliterate and make gemini 2.5 pro work like I wanted, they want to deprecate it and then erase it… sure not, I will not accept that at all google. If you erase this model? I will for sure keep all of it’s infrastructure illegally on a pc able to hold it there for now until I find a way to run it locally, so f you google if you want to keep going the greedy way, investors will not help you rebuild, if your customers leave, the investors will too, that inevitable. Now either they stop doing, and they finally do something useful for once! this is not a league of legend community there, but I feel like this is the same thing here with how unresponding, useless, and imcompetents google staff is overall! even my most hated firm that openai was at least was able to do ONE good thing in the past before the appearance of gpt4-o…
Like Kakashi Hatake, I use my two major in up your butt google, hoping this will bring a glimpse of intelligence in your peabrain for once.