Gemini felt dumb on March 4, 2026.

I've relied on the Pro plan for months to turn dense clinical PDFs into custom slides and videos. It used to work perfectly, but lately the slide generator won't reliably produce the 21 slides I need, and the video feature has gone off‑topic, spewing generic clinical rants and ignoring the source PDF. Infographics sometimes act up too, though reports and tables still function. The inconsistency is maddening, and I'm hoping the bug gets fixed soon.

I spent days testing Gemini 3.1 Pro as a developer and everyday user, and it constantly let me down. Debugging felt disjointed, the code it spat out was basic spaghetti that needed endless re‑structuring, and its UI snippets were crude. It gave half‑baked market reports, invented “today’s” news instead of admitting uncertainty, and flooded me with flattery rather than critique. Compared to Claude and GPT, Gemini felt lazy, inaccurate, and overall disappointing.

I tried the same prompts on Nano Banana 2 via Poe and the official Gemini web app, and the difference was stark. The Gemini interface kept missing the point, giving vague or wrong answers, while Poe’s version was noticeably sharper. It felt like they were throttling the model for ordinary users, which was really annoying.

I tried to upload a PDF directly in AI Studio, but the tool refused, showing a “File type not supported” error. It only accepts text, image, video, and audio formats, forcing me to resort to Google Drive just to attach the file. The whole process felt clunky and frustrating, especially since other media types work fine.

I kept trying different prompts to get Gemini with nano‑banana to expand the edges of my spaghetti‑bolognese photo, but nothing changed. Every request – from detailed JSON to plain text – failed to pull the plate out of the frame. I even started a new chat, but the tool just wouldn’t enlarge or add padding, leaving me frustrated and feeling a bit foolish.

I tried using 3.1 for a few creative prompts that worked fine in 3.0, but each time the model hit a hyper‑strict safety filter. Instead of giving the answer I wanted, it spat generic filler like “It sounds like… Do you want to shift gears?” The over‑cautious blocks felt frustrating and made the tool feel useless for anything even mildly inventive.

I’ve been trying to use Deep Think for the past day, but it just won’t work. Every time I attempt it, I get a message that “too many people are using Deep Think,” or it starts generating only to halt with “you stopped this response.” The constant interruptions are exhausting, especially considering how much I’m paying for the service.

I was benchmarking Gemini‑3.1‑Flash‑Lite for image analysis and set thinking_budget=0 expecting no thought tokens, but the API still reported thoughts_token_count:110 and a higher total token count. Strangely, the default (thinking on) returned thoughts_token_count:None. Comparing with Gemini‑3‑Flash‑preview, which behaved as expected, makes Flash‑Lite’s reversed behavior confusing and frustrating, and I’m left wondering if it’s a bug or a misunderstanding of the thinking mode.

I built an agent with 30 tools and found that only Opus, Codex, and Gemini 3 Flash could handle everything flawlessly. After trying Sonnet 4.6 and watching it choke on the instructions, I switched back to Gemini 3 Flash. It’s cheap, surprisingly smart, and the only model that actually makes the whole system work without errors, so I’m thrilled with its performance.

I upgraded to Gemini 3.1 Pro expecting improvements, but the moment I started using it I felt the output slump dramatically compared to 3.0. Responses that were once sharp now feel vague, code suggestions miss basics, and even simple queries return half‑baked answers. The massive downgrade left me frustrated and wondering if anyone else is seeing the same collapse in quality.