Gemini felt dumb on January 8, 2026.

I tried using Gemini 3.0 Pro on the web app and it felt crippled—responses were cut short, the session limit of 20‑25 prompts made it feel like a 100k‑token model instead of the promised 1M. The throttled rates and weird behavior left me frustrated, especially compared to Claude 4.5 Sonnet, which feels far more reliable. Now I’m stuck deciding whether to stick with Claude Pro and give up Google’s storage perks or wait for Google to sort out the mess.

I spent three months testing Gemini Pro 3 and kept hitting roadblocks. It misread numbers from photos, swapped answers, and ignored prompts to only do X and Y, doing everything instead. When I asked it to trim a PDF to 30 pages, it claimed more pages and gave me just one. Even precise research requests spilled over to unrelated topics. Compared to ChatGPT or Sonnet 3.5, it felt clumsy and unreliable, turning simple test questions into frustrating mistakes.

I tried using Gemini to help pick ski gear by sending it ten images of ads. It read some descriptions but completely messed up the rest—got half the brands wrong, ignored prices and locations, and even thanked me for only five photos. After several precise prompts it still failed until I pasted the info myself, leaving me frustrated and wondering if I need a fresh chat each time.

I tried Kling AI by feeding it my kid’s photo and watched it magically turn into the viral dance clip. The motion‑copying was spot‑on, and I was pleasantly surprised by how smoothly the animation synced. I also dug into the pricing and found the subscription surprisingly cheap, which made the whole experience feel like a win‑win. The whole process was fun and left me confident in the tool’s capabilities.

I tried to get Gemini Pro to create three images, even writing the prompts myself after it suggested them for my project. One prompt included the word “nude” (non‑sexual), but the model instantly blocked all generation with a generic “cannot generate” notice. Removing that prompt didn’t help—first it claimed to have made the images, then nothing appeared, and finally it only offered textual descriptions. The whole experience felt stubborn and frustrating.

I’ve been trying to use Gemini for days, and it keeps crashing once I hit around 120,000–140,000 tokens, far short of the promised 1 million‑token window. It feels rude that Google touts that huge context size while the model just dies on me. The experience has been frustrating and disappointing, especially after the hype.