I’ve been trying out Gemini 3.5 Flash in AI Studio’s Build section, and I have to say the results are horrible.
I give it a simple prompt (e.g., “Make me a fun amazing high-quality pixel platformer. Have a title screen too.”), and it takes 800s to make it and there’s so many unnecessary things I didn’t even ask for, like a sound syntheziser or other things, it is extremely gptified.
I tried 3.5 Flash in Antigravity 2.0 and compared to AI Studio, it’s an incredible model. Actually does what I tell it to do and really good results.
I understand that 3.5 Flash was mainly designed and optimized for Antigravity, but it shouldn’t be this … of a model in AI Studio Build. Gemini 3 Flash Preview is lightyears better in AI Studio build.
I’m not sure if anyone else is experiencing this issue or if it’s just me. Feedback and anything is welcome.
You are definitely not alone in experiencing this. This kind of “over-engineering” or “gptified” behavior is a known side-effect when a raw model is placed inside complex application-building frameworks. Here is a breakdown of why Gemini 3.5 Flash behaves so differently across these platforms:
-
AI Studio ‘Build’ Agentic Scaffolding: The Build section in AI Studio doesn’t just pass your prompt directly to the model; it wraps it in heavy background system instructions and agentic loops designed to generate a “complete, production-ready” application. This forces the model to overthink, causing it to inject unrequested features (like sound synthesizers) and skyrocketing the generation time to 800 seconds.
-
Streamlined Pipeline in Antigravity 2.0: Antigravity 2.0 likely utilizes a highly optimized, direct pipeline specifically tuned for its environment, stripping away unnecessary backend prompt layers. This allows the raw efficiency and compliance of 3.5 Flash to shine without artificial bloat.
-
Model vs. Preview Behavior: Gemini 3 Flash Preview might feel “lightyears better” because preview models are often less entangled with the heavy post-processing and structuring templates of the Build environment, leading to a much more direct execution of your instructions.
Actionable Tip: To bypass this bloat in AI Studio, try using the standard Chat section rather than the Build section so you have full control over the system instructions. Alternatively, inject strict constraints into your prompt, such as: "Do not add any extra features or audio synthesizers. Keep the code minimal and focus strictly on the core gameplay and title screen."