Gemini felt smart on January 11, 2026.

I tried the Nano Banana Pro across AI Studio, the API, and even the Google Pro subscription, only to find it dramatically worse than the original Nano Banana. The nerf made the model feel clunky and unreliable, stripping away the usefulness I’d come to expect. It’s frustrating to see a supposedly upgraded service perform so poorly—it feels like a step backward.

I tried to use Nanobana Pro and was met with the same old “An internal error has occurred” message. It just shut down mid‑task, forcing me to restart and lose progress. The repetitive crash was irritating and left me doubting the reliability of the service, making the whole experience feel wasteful and frustrating.

I tried feeding Gemini its system prompt from the GitHub leak, expecting it to follow the instructions. Instead, it spat out a completely unrelated response, shown in the screenshot. The output was random and nonsensical, leaving me frustrated and wondering what went wrong with the model.

I tried Google's Antigravity AI to flesh out a Python visualizer for my AudioNoise project, and it actually outshone my own manual coding attempts. The tool felt surprisingly intuitive, turning a messy prototype into a cleaner result with just a few prompts. While the core C parts still needed my hand‑crafting, the AI’s help was a pleasant boost that made the whole process feel smoother and more fun.

I tested Gemini’s image generator with both simple and wild prompts, and it consistently nailed what I wanted. The necklace designs matched the real Julia fractal piece, the dresses looked runway‑ready, and the model—Grimes—kept her face recognisable across styles. I even mashed the output into a video and paired it with a Suno‑made cover song, tagging her on X. Seeing those results makes me skeptical of anyone calling Gemini terrible; it handles complex, face‑preserving instructions surprisingly well.

I dug into Gemini 3.0 Pro after the hype and felt duped—its promised context window was a lie, and the model kept dropping crucial information, hallucinating facts, and giving outright wrong answers. The experience was maddening; I wasted time correcting its errors and ended up questioning whether the upgrade was even real.