I absolutely love Gemini.
On January 1st this year, I upgraded to Ultra.
After nearly two weeks of use, I’m a bit disappointed.
In a nutshell, I believe Google hasn’t fully figured out what Gemini services Ultra members should enjoy.
First, my primary motivations for using Ultra (before noticing Antigravity’s generous Claude Opus 4.5 Thinking allowance) were threefold:
1. I love NotebookLM—it’s indispensable for my research and decision-making now.
2. I was curious about Gemini 3.0 Pro Preview DeepThink’s performance. It claims to deliver superior code and logical outcomes.
3. I anticipated even more impressive performance from Gemini 3.0.
For NotebookLM, there’s nothing more to say—it remains consistently impressive. I crave more book credits so I can outpace top researchers from leading companies in less time.
But regarding DeepThink and Gemini 3.0’s performance? It’s been particularly disappointing.
I purchased AI premium services, correct?
But:
1. When tested against my evaluation dataset, DeepThink’s error rate was on par with Gemini 3.0 in AI Studio—especially in language logic, legal recognition, and financial rule-making.
2. True, DeepThink doesn’t suffer from performance degradation.
3. Crucially—DeepThink lacks integration with many tools. This is deeply disappointing. It cannot read my GitHub repositories or analyze video content I provide. Even though my GitHub code repositories are under 5MB and my videos are under 50MB, it still fails to analyze them. Not only does it throw
Error 9
, but it also consumes my entire daily usage quota!
4. DeepThink cannot be invoked outside the Gemini app window. This drastically reduces my control over it! It’s truly disappointing. I tried reverse proxying DeepThink, but Google enforces strict restrictions on Gemini access through this channel.
5. As an Ultra member, Gemini’s performance and quotas don’t translate to AI Studio or other channels. In other words, Gemini remains the same across all platforms: performance throttles when the time limit hits, and quotas run out when they run out. This means I still have to pay for Vertex, and I still have to pay for my Cloud. So the question arises: why should I buy Ultra instead of just putting my money directly into Vertex?!
In practice, the tools most beneficial to my work are either NotebookLM (which is essentially a separate product now?) or Claude Opus 4.5 Thinking on Antigravity. Claude’s code performance remains excellent and reliable. Its attention to context and engineering mindset, coupled with its refusal to cut corners, still outperforms Gemini 3.0 in these areas.
It’s a bit of a shame: I subscribed to Google’s Ultra plan, yet Claude is what satisfies me most.
I’ll renew my subscription next month and plan to keep it active through June this year. I’ll continue engaging with the community as I’ve previously stated.