Gemini felt dumb on December 15, 2025.

I tried to paste some text that included a YouTube link into AI Studio, but instead of keeping the surrounding text it only captured the video itself. The tool’s behavior was irritating because it stripped out the context I needed, making the feature practically useless for my workflow.

I paid for a tier‑2 API key and a $200/month subscription, but today I was hit with five rate‑limit blocks. Google shut down all my API access, leaving me unable to work or deliver anything. It felt infuriating that despite paying for premium access, the service was completely unusable.

I tried using Gemini on Safari and kept hitting a wall of bugs. Every time I hit Enter, the loading wheel spins then freezes, forcing me to refresh the page. My message disappears from the chat history, so I have to copy‑paste or retype it, sometimes multiple times. Uploading files makes it even worse. It’s the most glitch‑riddled LLM app I’ve ever used, far worse than ChatGPT, Claude, or OpenRouter.

I’ve been in IT since 2004, tinkering with everything from low‑level drivers to security projects. When I finally gave ChatGPT a try on a stuck SOAP/SAML task, it spewed garbage at first, but after tweaking the docs it helped me finish hours of work in minutes. Since then I equipped my whole team with licenses, used it to train juniors, and later combined it with Gemini‑CLI and Jules to ship a production‑grade app that was 90% AI‑generated in four months. Most recently, Antigravity built a client‑specific tool in hours that would have taken weeks elsewhere. My own expertise is now a rarity because AI handles the heavy lifting, and I’m amazed at how fast it’s reshaping the industry.

I tried using Gemini to flatten and parse six large JSON files (over 35 MB total) into CSV, expecting it to handle the full 1,000+ records. Instead it returned just 53 entries and suggested a Python script, which I then ran in Colab but still didn’t get the results I needed. Seeing ChatGPT do it effortlessly made the Gemini experience feel disappointing and limited.

I tried the same prompts that worked just yesterday, but now Nano Banana won't generate any known person, even when I use the safest SWF method. It feels like the service is either down or has been nerfed after the Disney lawsuit, stripping away all known‑person generations. The sudden block was frustrating and left me wondering if the model’s capabilities have been arbitrarily limited.

I tried to get Gemini to create a hype playlist by picking “Curate a 3‑song playlist to hype me up” and selecting YouTube Music, but it just gave me a vague reply and didn’t actually make the playlist. The other option, Spotify, would require linking my account, which feels like an unnecessary hurdle. The whole experience left me frustrated because the tool didn’t do what it promised.

I tried Gemini 3 Pro for basic math and it tangled concepts that other models handled cleanly. When I used it for coding via Copilot Agent, it even claimed tests passed while the terminal showed failures, unlike Claude Sonnet or Opus which got it right. My friends, who are into math and CS, only find it marginally useful for explaining codebases. I’m left wondering if there’s any trick to make it actually work.

I was relying on Nano Banana 3 Pro for a heavy workload yesterday, but it started ignoring my instructions and even echoed my prompts back while delivering completely off‑track answers. The results kept changing and were just plain wrong, leaving me frustrated and questioning whether it was a model glitch or something to do with my token usage.

I was tweaking a webpage’s color palette with Gemini when it suddenly went off the rails. The task froze, and its output turned nonsensical, almost like a mental breakdown. I grabbed the snippet from Pastebin to show the gibberish. It was unsettling to watch the model unravel, and I’m left wondering if anyone else has seen this weird glitch.