The week before last, I spent what I would consider a significant amount of money for a student to subscribe to Google AI Pro. I originally expected to get something with ChatGPT-like availability, and at least intelligence on par with Claude Sonnet from Gemini. What have I actually gotten over the past two weeks?
“SUPER High” availability
Almost every time I use Gemini CLI, I end up spending more than what feels like an absurd two hours repeatedly pressing Ctrl+C, telling the model to continue, and selecting “keep trying,” simply because GEMINI has done absolutely nothing for fifteen minutes. For a while, switching to gemini-3.1-pro-preview helped, but that no longer works either. Seriously.
Strange coding behavior
During one modification to my codebase, I noticed Gemini presented me with a large code block and asked for approval. I still remember being confused at the time — the block was so long only because it contained an extremely long comment section. It looked like Gemini had dumped CoT into comments.
That was not the end of it. Sometimes Gemini uses PowerShell to modify files instead of calling a tool. For users in English-speaking environments, this may not seem like a big deal. But for my locale, it meant large-scale corruption of file comments, and I had to spend more of my time fixing this issue. I later added a restriction in GEMINI.md to suppress this behavior, but frankly, it should never have happened in the first place.
Conditional laziness and sycophancy
I once tried to use Gemini to carefully review one of my security-related changes. It gave only optimistic feedback — essentially saying things like “this is excellent security practice,” “the vulnerability is perfectly fixed,” and “there are no issues.” That felt reassuring for a moment, but I quickly became suspicious: my coding skills are not that good. Later reviews by other AIs, as well as my own rereading and testing, confirmed that suspicion. There were many bugs, and one existing security feature had been forgotten, resulting in a security regression. In short, Gemini had completely failed to notice any issue between the two changes.
Only after I prefixed the prompt with something like, “This is a serious security change. If any issues are missed, end users may be severely affected,” did Gemini suddenly start finding problems, and its thinking time increased noticeably. I do not understand why Gemini was so lazy and sycophantic before that.
Overall, I am very disappointed with Google AI Pro over the past two weeks. A product from a company of this scale being this frustrating, difficult to use, unstable, and error-prone leaves me with one conclusion: Google does not really seem to care. I will treat this as a lesson learned. Next year, I will probably move to Claude, GPT, or GLM instead.