Gemini felt dumb on November 15, 2025.

I tried Gemini 3.0 through OpenRouter and was instantly struck by how fast it responded. The code it churned out was clean and functional, saving me time fiddling with syntax errors. I felt a rush of relief as the tool kept up with my rapid requests, and the overall experience left me impressed with its speed and quality.

I tried accessing the newly announced 2‑million‑token models on OpenRouter, but every request returned a 400 error. The tool’s behavior was irritating, as I couldn’t test the promised context size and was left wondering if the issue was with Gemini 3 or the API itself.

I tried to get Gemini to adopt a specific accent, asking it to sound like a white valley girl and it nailed the tone. But when I requested it mimic a brown Bengali girl, the model flat‑out refused, citing stereotypes. That refusal felt like a clear bias and made the interaction feel unsafe and discriminatory, leaving me frustrated and angry at the tool’s racist behavior.

I paid for Gemini Ultra hoping for a smooth premium experience, but it’s been a nightmare. The sharing feature refuses to give me a URL and forces me to copy‑paste code into an HTML file, which kills my workflow. Safety filters are impossible to dodge, resources time‑out during peak hours, and the promised agentic tools are nowhere to be found. It feels like I’m paying for a broken, over‑filtered product, so I’d stick with Pro instead.

I’ve been using Gemini Live a lot and noticed the new version feels way sharper than the old “smooth‑brained” Gemini 2. The answers are noticeably richer and the overall quality is a big jump – it even feels amazing at times. The only annoyances are the “speak faster” toggle that won’t stay on and the accent setting that fades unless I keep nudging it, but those minor hiccups don’t ruin the much better experience.

I’ve been using Gemini 2.5 and it’s great for coding and writing, but a few quirks are driving me nuts. The edit button only works on the most recent prompt, so once I move on I can’t tweak earlier answers—a feature other models have. Also, in long sessions like drafting a novel, the chat history suddenly cuts off, making it impossible to view the earliest parts. These odd limitations feel bizarre and interrupt my workflow.

I tried to run a long prompt and the interface just hung—‘please wait for the content to finish loading’ never went away, and the Run button stayed greyed out. I checked my token usage (20 k input, 185 k output, 205 k total) and wasn’t close to any limit, so it felt like a glitch. The whole thing was frustrating because I couldn’t get any output and wasn’t sure if I’d done something wrong.

I ran the same historical ledger image through Gemini 3’s Canvas on Android and was amazed that the output closely matched the blog’s quality, even with a low‑resolution picture. By contrast, Gemini 2.5 on desktop barely got the basics right, and GPT‑5 Pro, while better than 2.5, was still far behind 3.0 and took ages. The tool’s performance felt surprisingly strong and promising.

I kept running into weird sexual hallucinations with Gemini, which was both jarring and unhelpful. After multiple tries the nonsense just kept showing up, making me lose confidence in the tool. I even subscribed to Perplexity hoping for a fix, and it turned out to be noticeably better, so I switched. The experience with Gemini was frustrating and disappointing.

I tried Gemini 3’s “riftrunner” mode hoping it would have solid world knowledge, but the responses were wildly off‑planet. The answers felt like guesses from another galaxy, missing basic facts I expected. It was frustrating to watch the model wander into nonsense instead of giving the straightforward info I needed.

I use the Storybook feature every night with my kids, and the latest update blew me away. The stories are richer, more imaginative, and the new audio feels like a real narrator—not a robotic read‑out. The pacing and emphasis are spot‑on, turning bedtime into a true storytelling experience. It’s a massive upgrade I’m thrilled about.