I’ve noticed that Gemini Flash 3.5/3.6—even on the “High” setting of the “Antigravity” version—gets basic things wrong compared to Gemini Pro 3.1. I don’t know if they tried to copy the style of DeepSeek V4 Flash, but honestly, it’s behaving in a very limited way; it makes mistakes on simple tasks. I’m not talking about complex programming or massive game projects, just simple things. It’s perfectly natural for AI models to make mistakes (if we do, imagine them), but compared to other models—and even the 3.1 base model—this one leaves a lot to be desired. I know Flash models tend to underperform, but this one is really bad. I understand Google had to comply with safety restrictions before the 3.6 Pro launch, but 3.6 Flash is just too weak, and the whole “High/Medium/Low” model selection scheme is a hassle. Why don’t you implement an adaptive system like Anthropic’s, or give us a slider for token settings or a simple High/Medium/Low selector?
You created the Transformer architecture and possess the world’s largest scientific foundation in technology—including AI—yet you are falling behind. It’s frustrating for those of us who use Google products for development. This is honest criticism, and I hope Gemini 3.6 Pro proves to be far superior. I won’t compare it to models like Fable, Mythos, GLM, Kim K 2, Sun,Mini max 3,Qwen 3.8 etc., since every company has its own policies, but it’s hard not to notice the difference in model performance. It might not be an issue for casual conversation, but when it comes to programming, the situation becomes stressful. That’s my honest opinion; yes, I know there are other options like Codex, Code, Open Code—I study LLM programming—but Google needs to wake up.
I find myself both agreeing and disagreeing with you. Having frequently approached programming through the Antigravity IDE—which is rather chaotic and demands strict discipline—as well as the Gemini CLI and several other platforms, I have observed that while these models are remarkably swift, they do occasionally omit or corrupt code. However, when meticulously guided and closely supervised, they execute tasks flawlessly according to plan. Perhaps there is a need for more documented use cases and shared experiences. I recognize how frustrating this can be, yet it is entirely manageable. Projects must be developed incrementally, ensuring that functions do not intertwine, tests are continually refined, and the codebase remains pristine.
I agree with my colleague’s analysis: this model is clearly inferior to previous Gemini iterations. I can’t speak for other AI models since I’m not as familiar with them, but code generation has definitely taken a step backward. The model struggles with basic reasoning and just fails to grasp instructions
100% agree i put out a youtube video about this yesterday. The biggest worry is that when I want to make a major architectural change and as part of my rules it has to give me options. When I show Claude or Perplexity the options 100% of the time they have told me not to use any of the options. The most shocking example was today when antigravity sa8d I couldn’t use flash 3.6 as an endpoint. And wanted my to revert back to flash 2.5. I knew this was wrong and Claude confirmed it. I pasted Claudes reasoning and then antigravity apologies and agreed. Bonkers
Ok, nvm, there is something weird with Gemini 3.6 Flash; it is super inconsistent. Sometimes it works great, but there are times where it skips whole steps. I think they optimized it too much. It feels exactly like when you work with local models and they start to lie to you.
Edit: In this example they marked all their tasks as fulfilled. I asked back where the feature was, then in proceded to code it in place and give me it as the response instead on acknowledging it didn’t implement it. This is a silly project, but imagine this thing happening in a serious project. Good stuff I double check everything.
3.1 ask for more tokens but he tends to respond better and when he makes a mistake he notices and tries to correct but he doesn’t insist on continuing to make mistakes, 3.6 it seems the context is fixed in his mind (his attention is kind of …),It goes into a silly loop, it seems to me that they tried or trained using the flash of Deepseek v4, but that it got really bad… That training is kind of left aside so as not to say that they didn’t do anything…after some key heads left ALPHABET things don’t seem to evolve before what is happening in the sector and in the AI race…Google is lagging behind with this : “I m big and i launch a better one later!”
EXACTLY, it is very … in simple language, the product is really weak, it looks like a distillation of FLASH V4, you know, because I know that because I use opencode with v4 flash for free via openrouter and opencode zen and the same way of acting and thinking I can’t say that it is improved Distill and even if it forces distillation it is a common practice the only ones who cry about it Openai and Anthropic with this … of intellectual property when all big techs do this absolutely.Since you’re going to distill use kim k 3 hahahah at least we’ll have a high end model use the base model 3.1 and the kim k 3+GLM ready recipe to surpass everyone since you’re going to make copies that make it with excellence
I understand that some people have had bad results from Gemini 3.6 Flash.
But it has been working pretty well for me.
I am writing a music composition app in Kotlin/Compose. I am coding in small steps like adding a module, fixing a bug, etc. The code looks pretty good. I have not noticed any regression.
I am also using Opus 5 but it seems slower and keeps pestering me with requests to run gradle or grep. Gemini is able to work more without human intervention.
Ok, update. I think I have wrapped my head around this version, and I think I like it in a way. You need to hand-hold it, and the best way to do it is via rules. A feature/problem check/audit/assessment artifact plus Conductor tracks help a lot with that.
The only remaining problem I had was with the progressive disclosure of skills and explicit skill loading not working correctly. I ended up adding this rule globally:
## Skill Rules
- **Skill loading**: If user references skills then you **MUST** load them. User can prevent this behavior if they mention verbatim `prevent_skill_loading`.
Agreed. I mainly use Gemini Pro 3.1 over Flash 3.5/3.6. Claude Sonnet and Opus are superrior to Flash but their tokens run out in a single request starting from implementation plan to code changes for adding a feature/
True, I’ve been using local models for small projects on Atomic chat+turbo quant ,atomic..chat , Release Atomic Chat v2.0.5 · AtomicBot-ai/Atomic-Chat · GitHub ,runs on any local machine be it linux ,windows, mac,android and iphone ,with cuda ,without cuda or VulKan,suport ROCm , this helps me avoid using tokens,runs well on an rtx 3060 12GB in gguf Bansai 27, Gemma 4 12b, 26A4, Qwen3.6 27 and 31, 35A3 and other modified models,not geniuses, but they can perform some basic tasks when well-guided.
