Gemini felt dumb on February 22, 2026.

I hooked Gemini 3.1 up to the Codex CLI while building a 3‑D game, and it was a revelation. The model followed my prompts flawlessly and even anticipated the next step, making the canvas workflow feel seamless. I’m thrilled with how well the two tools mesh, though the tangled pricing options left me a bit uneasy and searching for a student discount.

I asked Gemini 3.1 Pro to write an original piece and then output ASCII tablature for it. To my surprise it actually generated a full, playable tab that I could copy‑paste and use. I’ve tried similar tricks with other models and they always fell short, so seeing this model nail it felt impressive and satisfying. The tool’s ability to compose and format the music left me genuinely pleased.

I spent hours digging through 15 PDFs—about 400 pages—to spot inconsistencies. ChatGPT kept tripping up: I had to batch‑upload, it hallucinated numbers, and forgot earlier docs. When I switched to Gemini I could dump all files at once, ask for a comparison table, and it instantly flagged three contradictions with exact page references. The workflow felt seamless and far more reliable.

I tried using GPT‑4o and Gemini as my personal tutor, asking them to explain topics, craft questions and clarify points I didn’t get. Instead, they kept missing key details, gave shallow answers, and I had to constantly correct them. It felt like a waste of time—watching a simple YouTube lesson would be far more reliable and stress‑free.

I ran a side‑by‑side test of Gemini 3.1 Pro and Claude 4.5 on a dense sci‑fi plot. Gemini kept the technical depth, stayed in first‑person, and never hit a safety block – it felt like a real collaborator. Claude, however, shut down, calling the story a “how‑to” for jailbreaks and refused to continue. The contrast was striking and left me both impressed and frustrated.

I keep hitting random “Content not permitted” blocks with version 3.1, and they show up in the middle of perfectly fine prompts for no reason. It’s unbearable—sometimes the model spits out a handful of lines, then counts that as part of my daily quota. It feels like the system is deliberately throttling me, making the whole experience frustrating and useless.

I’ve been watching Gemini “Fast” get worse over weeks, and it’s become a hassle. What started as a short, clear rule about dash spacing now needs extra sections just to function. I keep tweaking the instruction just to avoid mistakes, and the whole process feels both frustrating and oddly amusing as the model keeps forgetting the simple rule.

I’ve been hitting a snag where Gemini 3.1 Pro just quits “thinking” after only a few hundred tokens. I have to nudge it to keep going, but eventually it just shuts down and I have to start a new chat. I’ve seen similar complaints before, but now it’s happening even earlier. I’m using the API in AI Studio and wonder if anyone else sees this and knows why.

I tried using the new model for a few sessions and kept hitting roadblocks. It gobbled up RAM, my quota vanished quickly, and the agent died with errors far too often. Long contexts slowed everything down, sometimes even refusing to open. The UI gave me no clue about token usage, and the browser felt snail‑slow. Overall the experience was frustrating and far from what I expected.

I keep hitting a maddening inconsistency with Gemini when I ask for 2k images. Sometimes it nails the resolution, other times it drops to 1k even though I’m using the same prompt and conversation. The only reliable trick is to start the chat by requesting 2k; otherwise it feels random. Re‑uploading the same image sometimes fixes it, but the unpredictability is really frustrating.

I kept seeing AI overviews dismiss Bomellida as fake, even though I have proof it exists. The tools confidently hallucinated, treating lack of Wikipedia data as evidence of non‑existence. It was frustrating to watch generative layers fill gaps with wrong assumptions, forcing me to argue with the AI that it was simply wrong.