I wouldn’t have faith in 3.5 Pro given how they have been internally filtering 3.5 Flash’s model. I can do all sorts of stuff with say 3.1 Pro that 3.5 Flash just strait up denies in response. But fingers crossed.
Beyond that though, I’ve been mulling all this over and could use a hand figuring out if I’m just going nuts or if there’s something to this below. Maybe it builds off a bit what you already brought up.
I’ve been following the discussions here about the recent safety updates in AI Studio—specifically the shift from soft checks to “full wipes” and the general frustration with the system being over-sensitive.
I’ve been looking into the underlying mechanics of how these updates affect us, especially those using paid subscriptions, and I wanted to run a few thoughts by the community. I’m curious if you are seeing the same patterns, or if there’s something I’m missing here.
There seem to be three major contradictions in how the platform is currently running:
The Paid Quota Drain (Paying for Blank Screens):
For those of us using premium plans like Google AI Ultra or AI Pro via Google One, we aren’t using pay-per-token APIs. We pay a flat monthly fee for a prioritized, daily/hourly prompt quota.
But have you noticed what happens when a response gets blocked/wiped by the system above the model itself after generating? It still consumes your paid daily prompt quota.
In standard consumer transactions, paying for a service, having the platform process it, and then completely erasing the final product before delivery while still charging you for the attempt feels incredibly wrong. If they block it, we shouldn’t lose the paid query. Are you guys seeing your daily limits drop after a wipe?
The Non-Transparent “Gambling” Loop:
Normally, when software or an API fails, it gives you a diagnostic error (like a copyright block, or a specific safety category) so you can fix your prompt.
Now, because the wipes are a complete black box and generally with zero feedback, we have to blindly guess what triggered the filter. This basically turns prompt writing into a game of chance:
-
We risk a finite, paid resource (our daily prompt quota).
-
The outcome is governed by black box, highly volatile post-filter with seemingly zero context awareness.
-
If we lose, our quota is gone, we get a blank screen, and we have no info to help us fix it for the next run beyond dumping more generation’s and hoping for a result.
Does anyone else feel like they’re just pulling a lever and hoping for the best now?
The Visual Design Paradox:
There’s been some pushback from product managers saying “AI Studio isn’t a creative writing tool, it’s a developer prototyping platform.” But does the physical design of the UI actually reflect that?
Think about it:
-
The UI is built like the ultimate writing playground: It has a massive dynamic up to million context window for the flagship models (perfect for loading entire novels or writing your own), a System Instructions box (perfect for character personas), and Temperature/Top-P sliders (specifically for controlling creative prose variability).
-
Actual production devs don’t use a manual chat box: For real enterprise scaling, developers use programmatic tools, Python SDKs, and automated testing. They generally from my experience aren’t sitting there typing back-and-forth in a chat window like the one in Google AI Studio’s playground.
By sanitizing this manual chat playground, they are breaking a very good creative writing engine outside of ones specifically built for creative writing on the market to appeal to a corporate crowd that it doesn’t seem to have the tools compared to say Antigravity or Vertex and jazz.
What do you all think? Are these observations aligning with what you are experiencing on your end? I wanted to put this out here first to see if we can verify these points together cos I ain’t exactly a power user outside of creative writing.
