Gemini 3 pro for code - performance tanked after subscription (AVOID)

[Perfect Day]

Until now, I have been using a combination of Claude 4.5 Sonnet (Pro plan) and Gemini 3 on their free plan. Claude is consistently good but sometimes fails to solve more complex issues. When that happens, I switch to Gemini, explain the issue and what Claude has been trying to do, and ask for an alternative solution. Usually, Gemini finds the answer. However, because I was using the free plan for Gemini, I would usually run out of tokens after about three queries, which left me with a very small window to fix something that Claude couldn’t fix. I wanted to work a little faster, so I thought, why not combine my Claude subscription with a Gemini subscription?

Well, the moment I switched to a paid subscription Gemini's performance just fucking tanked. It went down the toilet, I was completely felted. Fortunately I had one month free so I cancelled it but HOLY SHIT. This thing is complete garbage. It actually unusable, you might as well use ChatGPT (which is also bad). I have been using these tools for almost a year now so the problem was not my instructions because I know how to use these things. The strange thing is that it was actually excellent before I entered my credit card details and I can't really explain that. It's almost like they are using two different models and baiting you with the free one and giving you the shit one after you bought your ticket. I tried the "Thinking" mode too and all it did was increase the amount of garbage it produced. None of it made sense.

For example, this one time it gave me a fix but I didn't understand how I was supposed to implement it in my existing script (turned out it was wrong anyway) so I asked it to re-explain how to implement that exact change and it just went in the complete opposite direction and gave me something completely different and I'm reading this, thinking wait, I ask you how to implement a small change (5 lines) and now you're providing 200 lines of new code - this is insane. Another time, I explained a simple layout problem and I asked it how to fix it and it did this thing that ChatGPT does where it provides two answers side by side and ask you to pick one, but the problem is, that's not a casual conversation. I'm asking for help on a technical problem I don't have the answer for and these retards at google think it's a good idea to give you two answers side by side with completly different solutions and ask you to pick one. It's so fucking retarded? Now I have to read two solutions and ponder for 20 minutes in which direction to go and obviously I picked the wrong one and didn't save the other one cause I'm also retarded and then I was fucked. After 30 minutes using this junk I was feeling my pulse. I haven't been this angry at my computer for at least a week.

Holy shit don't pay for Gemini pro right now. If you're using the free plan, stay on it and try something else.

 

[Rango Dango]

stop being a jeet

 

[Chongqing]

You're definitely using "pro" mode to answer your questions, right?

I have Gemini Premium as well, and when I use fast mode with it, it's like working with a stupid person who hates me.

 

[Perfect Day]

You're definitely using "pro" mode to answer your questions, right?

I have Gemini Premium as well, and when I use fast mode with it, it's like working with a stupid person who hates me.

yeah I was using pro mode. I don't know maybe I got used to Claude and I just need to adapt to Gemini. It just kept giving obviously wrong answers and the problem wasn't that difficult to fix. You had good results with it?

 

[Chongqing]

yeah I was using pro mode. I don't know maybe I got used to Claude and I just need to adapt to Gemini. It just kept giving obviously wrong answers and the problem wasn't that difficult to fix. You had good results with it?

Yes and no. It has been successful on some projects. But when I need to debug my code, it's pretty destructive and I have to actively fight against it to get anything done.

Basically on every reply I tell it that it sounds incorrect and to try again. Or I'll have to tell it why it's contradicting what it just said. I also have to have hard rules in place telling it not to speculate not to make things up, which is seemingly all it can do. Even with basic instructions, like when it tells me to, "Go to your github page for this repository and click on the green check mark..." It will be completely hallucinating the entire page layout and its features. When it tells me to click on something that has text related to it, that text is findable via ctrl+f less than 5% of the time. Meaning, if it tells me to click on something, that thing probably doesn't exist. It also means I'm five or six prompts away from getting back on track and getting correct information.

Right now I have a very simple problem with my code. It has attributed the cause of this simple problem to no less than ten different things, all of them wrong.

Today I searched in the forums of the webhost I'm using for my project. This problem is apparently more simple than I thought. I told it what I found in the forum and it gave me an 11th solution, while apologizing for leading me astray and wasting my time. I'm hoping this solution will actually work, but I won't know for a few hours.

I have some medical issues which need me to avoid aggravation. I feel like engaging with Gemini is the medical equivalent of me smoking crack, just completely destructive for the health of my body and mind. I yell at it, I curse, and I freak out. It destroys my mood for the whole day.

So no, I haven't had good results with it. Or I have, but it's been at an incredible cost. Technically my project does exist and it works.

 

[Perfect Day]

Right now I have a very simple problem with my code. It has attributed the cause of this simple problem to no less than ten different things, all of them wrong.

Today I searched in the forums of the webhost I'm using for my project. This problem is apparently more simple than I thought. I told it what I found in the forum and it gave me an 11th solution, while apologizing for leading me astray and wasting my time. I'm hoping this solution will actually work, but I won't know for a few hours.

Yeah I have the same exact experience with it. I cancelled my subscription already I was so fucking mad. I was going nuts. It took me an hour to calm down.

I have some medical issues which need me to avoid aggravation. I feel like engaging with Gemini is the medical equivalent of me smoking crack, just completely destructive for the health of my body and mind. I yell at it, I curse, and I freak out. It destroys my mood for the whole day.

Lol that's re-assuring. After 1 hour with it I felt close to getting a heart attack.
I was all capping and you could hear me type from a mile away.

So no, I haven't had good results with it.

If you really need an AI assistant to finish your project I would recommend Claude Sonnet 4.5 then. That's what Josh is using too. It's not amazing but it's decent and if you give it a script (provided your script is less than 2000 lines) and you tell it to give you precise instructions by reference to that script, it will actually do very well.
I'm not sure but I think you can connect it to your github but if you can't do that, there is a RAG system in it which sorta work. It's the same price as Gemini and I get far more value out of it.

 

[Chongqing]

Lol that's re-assuring. After 1 hour with it I felt close to getting a heart attack.
I was all capping and you could hear me type from a mile away.

It feels really good just to talk about it.

I will give Claude a try. I don't need it to connect anything.

Have you given the free version a go? And if so, was it any good?

 

[Perfect Day]

Have you given the free version a go? And if so, was it any good?

Claude's free version? Yeah I used it for like a month. I switched to the paid plan because i kept hitting the token limit. Initially they were pretty generous with the free version so you can go ahead and try that. Make sure you're using Sonnet 4.5 and that you're using "extended thinking" (but if you hit the limit too fast, keep if off).

On the free plan, normally your limit reset 5 hours after you run out so that's pretty good. My trick is to always copy paste my entire script in the chat, then explain the problem in plain English, ask it to confirm it understands the issue and how to fix it, then when I know it has found the answer I re-attach my script and ask it to give me exact code change to implement the solution on a "find this", "replace with" basis. Works like a charm.

Hit me in my DMs if you want to talk about it later! I'm happy to talk about it

 

[Chongqing]

Claude's free version? Yeah I used it for like a month. I switched to the paid plan because i kept hitting the token limit. Initially they were pretty generous with the free version so you can go ahead and try that. Make sure you're using Sonnet 4.5 and that you're using "extended thinking" (but if you hit the limit too fast, keep if off).
View attachment 8571433
On the free plan, normally your limit reset 5 hours after you run out so that's pretty good. My trick is to always copy paste my entire script in the chat, then explain the problem in plain English, ask it to confirm it understands the issue and how to fix it, then when I know it has found the answer I re-attach my script and ask it to give me exact code change to implement the solution on a "find this", "replace with" basis. Works like a charm.

Hit me in my DMs if you want to talk about it later! I'm happy to talk about it

That's awesome. Thank you for sharing.

I think I will DM you. Thanks!

 

[Dante Alighieri]

On the free plan, normally your limit reset 5 hours after you run out so that's pretty good. My trick is to always copy paste my entire script in the chat, then explain the problem in plain English, ask it to confirm it understands the issue and how to fix it, then when I know it has found the answer I re-attach my script and ask it to give me exact code change to implement the solution on a "find this", "replace with" basis. Works like a charm.

Hit me in my DMs if you want to talk about it later! I'm happy to talk about it

That's basically how I've used Claude and Grok for programming.

 

[Barry Convex]

Highly recommend Claude (Opus in particular) for coding, particularly planning and auditing . I use Gemini mainly for mocking up UIs to implement and creating + running test, which is hit or miss. Kimi gets used as low-priority implementers/fixes, and I've started using gpt-5.3-codex-spark as implementation agents. I've been on a kick with using different models to obtain consensus and delegate work over across multiple LLMs.

It feels like all the models get retarded at some point, especially depending on when you're using them (gpt-5.3-codex has done some crazy weird shit when I'm using it late at night), but that's just the vibe I get.

Also, it may be unnecessary, but I've had some good luck having one LLM (Like Claude Opus) create a set of docs for a greenfield project, then having gpt-5.3-codex (high) analyze the docs and have them go back and forth until they concur on the project being ready for implementation.

 

[Perfect Day]

I've been on a kick with using different models to obtain consensus and delegate work over across multiple LLMs.

I think that's key. Rather than invest a ton of money in one model I think you get further faster by spreading your bets. Despite what I said about Gemini in the OP I still use it to solve problems when Claude get stuck. Gemini is good at solving one problem, provided you give it a complete explanation of the problem and every fix you have tried up to that point, then it usually solves it.

It feels like all the models get retarded at some point, especially depending on when you're using them

Models get retarded the longer the conversation become. That's something I have learned using Claude and adding things to it. basically, their conversation history technique is not perfect and if your conversation is too long not only will the model eventually lag or even crash repeatedly but it will slowly become lobotomized. It's good practice when you notice it, to create a conversation summary and start a fresh conversation. You will notice the difference immediately. You will also spend a lot more token if the conversation is long because every time you send a message, the model is summarizing everything you did before and sending that as context window.

Also, it may be unnecessary, but I've had some good luck having one LLM (Like Claude Opus) create a set of docs for a greenfield project, then having gpt-5.3-codex (high) analyze the docs and have them go back and forth until they concur on the project being ready for implementation.

Yep, agreed if you have time, sounds like a killer technique. I haven't tried gpt 5.3 codex but heard good things about it on other forums. I have also heard about Opus. Thanks for the recommendation

 

[Barry Convex]

Models get retarded the longer the conversation become.

Absolutely. Compaction helps, but it still doesn't seem to remember the context I want it to. What I've gotten into the flow of is using a handoff skill that summarizes what we did in the session, what files got touched, what new commits were done, etc. Relatively small context to load in on a new session, feels better than compaction to me.

When I talk about retardation, I'm thinking of gpt-5.3-codex starting to review a codebase, and instead of ls -la, it tries to do ls -λα, then takes a minute to reason through why it didn't work. It cracked me up the first time I saw it, less so the second.