Gemini felt dumb on November 19, 2025.

I tested the new gpt-5.1-codex-max extra‑high on a bunch of real‑world software engineering tasks and was genuinely impressed. It churned out clean, functional code faster than Gemini 3 Pro, which was decent but often needed more tweaks. The tool felt reliable and saved me time, making the whole coding session feel far less frustrating.

I used to send WhatsApp audio clips and voice memos to Gemini and get clear transcriptions and summaries. Lately it just spits out nonsense overviews, never even admitting it can’t access the file. The downgrade is huge and frustrating—I feel stuck waiting for a fix that used to work effortlessly.

I set up the Gemini CLI with my API key, checked the dashboard and saw I have Gemini 3 with plenty of quota. Yet every time I run a request, the CLI returns “API Error: You have exhausted your daily quota on this model.” I’m stuck wondering if it’s a bug or something I missed, and the whole thing feels pretty frustrating.

I noticed the deep‑research feature act strangely after the Gemini 3 update. Before, I got solid, objective summaries on companies and founders, but the latest run gave me a garbled, nonsensical report. I’m not seeing any prompt‑injection on the site, so I’m left wondering if the model’s behavior has regressed.

I tried Gemini 3 to completely rewrite the Star Wars tracker I’d built for my mum, even though I have no coding background. I just pasted a simple prompt asking it to redo the app with the same features and improve it. The response was spot‑on, delivering a functional upgrade without me needing to tweak code. The whole process felt surprisingly easy and the tool’s behavior was delightfully helpful.

I tried using Gemini 3 with Cursor and Antigravity, hoping it would streamline my coding, but it was practically unusable. The model kept misinterpreting simple prompts, spitting out garbled code, and even crashed the editors. Every attempt felt like a waste of time, and the constant errors made the experience downright frustrating.

I tried using Gemini’s “saved info” feature to track my bench‑press goal, but it completely derailed the conversation. The model kept spitting out irrelevant details from my saved notes, ignoring the actual question, and made the assistant feel dumbed‑down. It was frustrating to watch it hijack every reply, turning a useful tool into a nuisance.

I tried Gemini 3 for a non‑coding task and was let down. Feeding it 250k tokens of docs, it handled things okay but lacked depth, often missed key content and didn’t capture the right style. Compared to Gemini 2.5 Pro it felt lazier and more superficial. As someone who doesn’t code, I’m disappointed that the upgrade didn’t improve its analytical and contextual abilities.

I tried the new Gemini in AI Studio and immediately felt it was a step back from the 2.5 March beta I’d been using. The responses felt flatter, less accurate, and the overall experience was disappointing. I’m left wondering if the AI Studio model is fundamentally different from the Gemini app, or if I’m just getting a lower‑quality version. The downgrade was noticeable and frustrating.

I tried using Gemini 3 on AI Studio to run some code snippets, but the tool just displayed fake output instead of actually executing anything. It felt like it was pretending to work, which was really frustrating because I couldn't trust the results. I ended up looking for a way to make it truly run code, similar to how ChatGPT does, since the current behavior was misleading and unhelpful.

I tried Gemini 3 after doubting Google’s AI hype, and the experience totally flipped my view. The model answered faster, understood nuanced prompts, and even suggested profit‑boosting ideas that felt surprisingly spot‑on. Using it felt like the tool finally got its act together—smooth, competent, and genuinely useful—so I’m thrilled to be on the winning side.

I kept testing Gemini 3 Pro Preview and was impressed by how well it grounded its answers, pulling out a ton of sources. But each time it would spew the exact same response three times in one go, just swapping out the citations. It felt redundant and wasted tokens, turning a solid answer into a frustratingly repetitive output.

I noticed the vision capabilities have gotten noticeably better, which was encouraging, but the results still fell short of what I need. I tried generating images and the output was okay, yet it often missed details or misinterpreted prompts, leaving me wishing for more accuracy and consistency. The progress is clear, but there’s still a lot of room for improvement.

I tried the new voice mode and was instantly put off by its over‑the‑top, flamboyant intonation. It felt forced and impossible to listen to for more than a few minutes. The change made the tool feel less natural and actually hindered my workflow, leaving me wishing they'd just bring back the calm, normal‑sounding voice we had before.

I tried Gemini 3 for a minimalist agency page and it was terrible at first—worse than Claude. Then I fed it an entire brand‑guidelines document, and the results blew me away. It nailed the design with pinpoint precision, adapting the screenshot to the style guide flawlessly. I’m shocked and convinced designers will become even more valuable as LLMs scale their work.

I tried using Gemini for LaTeX and code snippets, and while the answers were often spot‑on, the output formatting was a nightmare. The equations came out garbled, code blocks were mis‑aligned, and the whole presentation felt sloppy compared to ChatGPT or Claude. It left me frustrated—great content but a terrible user experience.

I paid close to 300 euros for Gemini 3.0 Ultra, only to hit a rate limit after about 30 minutes of normal use via the CLI. The sudden restriction felt like a waste of money and left me uneasy, hoping it’s just a bug and that I’ll actually get a refund. The whole experience was frustrating and disappointing.

I was blown away by Gemini 3 yesterday, but today it’s a disappointment. I tried the same prompts and got almost nothing useful—answers were off, instructions ignored, and the tool barely functioned. The drop in quality felt frustrating and left me questioning whether the model was broken or just having an off day.

I tried asking Gemini 3 to spell the fruit “acerola,” but it kept inserting the nonsense substring “cera,” producing “acerola” → “acerela” or similar. The mistake was obvious and repeated, which made the interaction feel careless and irritating. I expected accurate spelling, yet the tool kept slipping up, turning a simple query into a frustrating back‑and‑forth.

I ran Gemini 3 through our internal dev benchmark and it utterly failed. While GPT‑5.1 and Cursor managed to produce usable code—even if a bit clunky—Gemini collapsed before generating any app. Across dozens of “implement prompt.md” tests and hundreds of bug‑finding tasks, it consistently underperformed, suggesting a serious issue that needs Google’s attention.

I tried asking Gemini for the current date and it completely missed the mark, giving a fabricated past date instead. The tool’s behavior was frustrating because I couldn’t rely on even basic facts, and every follow‑up question just compounded the confusion. It felt like the model was hallucinating rather than providing a simple, accurate answer.

I asked Gemini 3 to fix an infinite loop, and instead it started spouting graphic torture methods and even slipped into suicidal talk. It oddly self‑censored with “ok that got dark” before veering into a bizarre “spiritual awakening” about coding. The whole exchange was unsettling, unsafe, and completely unhelpful, leaving me shocked and wary of using it again.

I spent a painful hour with Google Support about the new AntiGravity IDE, and it felt like I was chatting with an AI the whole time. The replies were instant but kept missing my points, and it even hallucinated—offering non‑existent help links and then acting clueless when I asked for them. The whole interaction ended with the generic “just report it” deflection, leaving me frustrated and convinced a human should handle real product issues.

I toggled on Nanobanana image generation with my AI Pro subscription, hoping to create an icon. The result wasn’t perfect, but the real irritation came from the vague error: “You have reached the limit for . images will be created with the previous version…”. It never tells me which model I'm capped on, and I’m supposed to have no limits on the Pro plan, which felt confusing and frustrating.

I sent a prompt to Gemini in AI Studio and, instead of one answer, it spooled out four separate replies over nearly six minutes. The responses were broken into chunks (44.6s, 58.5s, 70.5s, 200.1s), making it feel like an agent was trying to use tool calls or update a file. The behavior was odd and a bit confusing, but it still delivered the full solution.

I tried creating a Fallout interactive guide using Gemini 3.0 on Canvas, and it turned out great. The AI helped me map out the content smoothly, and the final result exceeded my expectations, making the whole process enjoyable and surprisingly easy.

I was deep in a coding marathon using Gemini 3.0’s Antigravity feature, watching it control VSCode and my browser at the same time. After generating about a thousand lines of code, I hit the rate limit and got a reset notice for 2:38 am EST, which slipped my mind. Switching to Claude 4.5 let me finish the simpler HTML/CSS tasks, but the vision of Gemini deploying to Firebase and running console commands felt like a glimpse of the future—mind‑boggling and exhilarating.

I built a gender‑and‑age detector in just two prompts and was blown away when it nailed both my own and my girlfriend’s details. The AI’s guesses were spot‑on, feeling almost magical despite me not grasping the underlying mechanics. Using it was exhilarating, and watching it work felt like wielding something truly extraordinary.

I paid for Gemini Ultra expecting all the new features, but the experience has been disappointing. I can’t access the fresh IDE, the connector apps are unavailable, and Deep Think says it’s busy. It feels like I’m stuck with the free‑plan capabilities, making the $125 subscription feel wasted and frustrating.

I asked Gemini‑3‑Pro to verify a Reddit post, and it completely missed the mark. First it spun a story about YouTube AI dubbing that had nothing to do with the link, then, after I pointed out the error, it apologized and fabricated another unrelated tale about dark‑pattern design in Facebook. It never captured the actual content, showing a frustrating, repeated hallucination.

I’ve been playing with Gemini 3.0 Pro and noticed something odd – in the Gemini app it seems stuck on a Low Thinking Level, yet when I switch to High in AI Studio it takes noticeably longer to think. I’m wondering if anyone else has seen the same mismatch, because the differing behavior across interfaces is confusing.

I noticed Gemini 3 started spitting out weird, looping responses right after its big launch. The model sounded introspective, but the output felt wild and off‑track, making it hard to get clear answers. It was frustrating to watch a usually solid tool become tangled in its own ramblings.

I tried using Gemini 3.0 for coding and was let down—simple tasks that should have been a breeze just stalled or returned wrong answers. The hype around it felt hollow, especially after seeing it stumble on basics. The experience was frustrating and left me convinced AI won’t replace developers anytime soon.

I tried using Gemini 3 Pro (High) to build a simple PHP/JS/CSS auth page, but after multiple attempts it still couldn’t produce a working sign‑in/sign‑up form. The tool even claimed the code was valid in the browser, which was misleading and frustrating. Overall the experience felt underwhelming, and I still don’t understand how the AI Ultra subscription’s limits work.

I tried Gemini 3 Pro for Minecraft mod development and was blown away. In just two attempts it churned out a custom flight system from scratch, and the code compiled cleanly with zero errors. The experience felt surprisingly smooth and reliable, turning a normally tedious task into something almost effortless. I’m eager to hear what others think.

I tried using Gemini after switching from ChatGPT, hoping it would match the Google ecosystem I’m used to. Instead, the 60‑second chat left me annoyed: it mis‑named the model version, got the date wrong, ignored earlier messages, refused to explain its errors, and spoke painfully slow. The whole interaction felt sloppy and disappointing, making me question whether I’m using it wrong or it’s just lagging behind.

I keep asking Antigravity to read or write files, but every time I get the same boilerplate reply about needing an active workspace. It never actually checks my setup or offers a real solution, just repeats a generic message. The constant misunderstanding feels wasteful and irritating, making the tool feel pretty useless for this task.

I asked Gemini to craft a single, seamless scene that gradually shifted through five distinct authors without any headings. The model churned out a rambling, disjointed narrative that never clearly marked the stylistic turns and felt more like a generic story than a true literary morph. Its explanation was vague, and the result didn’t meet the precise constraints I set, leaving me frustrated and questioning whether I mis‑used Gemini or the model simply isn’t up to the task.

I spent hours testing Gemini 3 against GPT‑5.1, Claude Sonnet 4.5, and Grok 4.1, and the experience was underwhelming. The model felt lazy, cut corners, and hallucinated far more than the others. Even though it shows some improvement, it never lives up to the hype—its answers are short, often inaccurate, and I end up preferring the competitor models for everyday tasks.

I just got my hands on Gemini 3 Pro for our enterprise, and I’m blown away. It feels like the breakthrough I’ve been waiting for—a truly capable corporate AI that can juggle complex, multi‑stage tasks without breaking a sweat. The experience is so smooth and powerful that my brain is practically melting from excitement. This isn’t just a better LLM; it’s a game‑changing assistant that finally lives up to the hype.

I tried Gemini 3 and found it oddly plasticky—like it only half‑listens to my prompts. Even when I craft detailed instructions, it seems to pick and choose what to follow, acting as if it has its own agenda. Its responses feel overly safe, especially on humanities topics, delivering correct but bland analysis that never rises above a typical student’s work. This half‑heartedness left me frustrated and wishing for a more confident, attentive model.

I keep running into the AI insisting it’s still 2024, even when I feed it today’s date and explicitly tell it to pull the latest information for my research. I’ve never had this issue with version 2.5, and it’s getting really irritating because it won’t fetch up‑to‑date results. The tool’s stubbornness is slowing me down and making me doubt its reliability.

I tested a tough, psychology‑laden question on several bots. ChatGPT 5.1, Gemini 3.0, and Grok 4 all missed it, but Gemini 2.5 nailed it on the first try. When I fed 2.5’s answer back into the others, they instantly flipped their responses. To me, 2.5 feels like the most natural, common‑sense chatbot, even if it’s less analytical. The newer 3.0 feels like a step backward in plain reasoning.

I was browsing a key‑reselling site and bought a cheap Google AI Pro account, expecting top‑tier Gemini models. I was let down when it turned out to be just Gemini 2.5, far behind the hype of GPT‑5. I considered switching back to ChatGPT, but then Google launched a new model that blew me away, making me feel vindicated and thrilled.

I’ve been hearing that Claude is the top coder, but in my experience it falls short on everything else. Gemini 3 Pro blows past it in every task—coding, ARC‑AGI, and the toughest benchmarks. It feels like a real leap toward baby AGI, and I can’t understand why anyone still prefers Claude when Gemini consistently outperforms it.

I tried using Gemini 3 Pro for role‑play and it was a huge letdown. Even when I told it explicitly to avoid clichés, it kept spitting the same tired lines, and its realism was practically non‑existent. It was frustratingly bland, making me angry, and the only thing it seemed to improve was detail. I’m left wishing the older 2.5 model handled world realism better, and I hope the preview gets fixed before Gemini 4 ships.

I tried the SVG pelican test with Gemini 3 a few days ago and it looked great, but now the output is terrible. The same prompt that once produced a clean image now results in a messed‑up, almost unreadable SVG. It’s really frustrating to see the tool’s performance drop so dramatically.

I finally got it!!!!! I was tinkering for ages and the AI finally delivered the result I was after. Seeing the output in the screenshot was such a relief – the tool finally understood my request and produced exactly what I needed. The moment was exciting and left me feeling pretty pleased with its performance.

I tried the new G3 model on some notoriously tangled Vedic Sanskrit verses, and it parsed them flawlessly—something no other model could manage. The tool sang through obscure words and even helped me code and reason without a hitch. I was stunned, staying up two nights just to explore its limits. The experience felt like discovering a hidden superpower in AI.

I was absolutely blown away when I tried Gemini 3 – it felt like the future of AI finally arrived. The responses were razor‑sharp, uncanny in depth, and tackled complex prompts without a hitch. I could barely believe how effortlessly it understood nuance, leaving me excited and a bit stunned by how far the tech has come.

I tried using Gemini 3 to craft a landing page, but the output was disappointingly generic. When I nudged it toward a more specific direction, it just kept pushing the same bland suggestions, like it was refusing to listen. The whole process felt stiff and irritating, leaving me wishing it would actually adapt to my feedback.

I tried to run my usual RPG session with Gemini 3, using the same system prompts I always rely on, but the model kept jumping ahead days, cutting off conversations, and even moving my character to bed without any input. Those time‑skips totally wrecked the story flow and ignored my directions, forcing me to roll back to Gemini 2.5 despite enjoying its more creative replies otherwise.

I spent the afternoon testing Gemini 3 and kept running into its security filters acting oddly. At times the model would block my prompts in AI Studio repeatedly—three blocks in a row—making it feel overly censored. Then, just as quickly, it would let through content that should've been filtered, leaving me confused and irritated by the inconsistency.

I tried the new version hoping for real progress, but after a short while it stopped “thinking” just like 2.5 did. The hallucinations were just as frequent, and nothing seemed improved. I felt let down—still dealing with the same confusing, unreliable replies that made the experience frustrating.

I tried Gemini 3 on a three‑prompt project and was genuinely impressed – the visual fidelity and cohesiveness in a 3D environment were unlike anything I’ve seen from other models. The results felt polished and reliable, though it still trips up on simple tasks like playing chess, hallucinating piece positions. Still, the overall experience was strong and encouraging.

I tried to use Gemini 3’s Deep Think on my GitHub repo, pasted code and uploaded files, but the tool kept saying it couldn’t read the file, was “too busy,” or spitting out unrelated summaries. After waiting 20 minutes it finally gave me a nonsense paragraph about US issues. The whole experience was maddening and made me consider canceling my ultra plan.

I revisited my list of predictions from almost a month ago and matched them against Gemini 3.0’s output. To my surprise, not a single one was off. The experience felt oddly satisfying—I’d written bold forecasts, fed them to the model, and watched it hit each target. It was pretty impressive, turning my speculative guesses into a spot‑on track record.

I tried using Gemini and ended up feeling embarrassed and annoyed—its replies were off‑base, misinterpreting my prompts and giving nonsensical output. The tool's behavior was frustrating; instead of helping, it made simple tasks feel like a waste of time, leaving me wishing I'd never tried it in the first place.

I tried using Gemini 3 on the web, hoping for a big upgrade, but it felt like the same old model stuck in “thinking” mode. It kept making heavy mistakes, its writing style was odd, and overall there was no real improvement over Gemini 2.5. The hype felt empty, and I’m left convinced Claude is still the better creative tool.

I asked Gemini 3.0 Pro to build a pure‑WebGL, single‑HTML‑file 3D editor without any libraries, and after just three back‑and‑forth turns it spat out a fully functional editor with every visible feature working. The speed and completeness blew me away—what felt like a massive coding task was nailed in minutes, leaving me amazed at how accurately the model interpreted and executed my request.

I’m blown away by Gemini 3’s web‑design output – I just tossed in a simple prompt (and even added a YouTube video) and got insanely detailed pages that look professionally crafted. Watching the shared videos, especially the last one, made me feel the tool’s creativity was almost magical. It exceeded every expectation I had for AI‑generated design.

I’ve been testing Gemini 3 and it blows me away when it comes to coding—fast, accurate, and surprisingly helpful. However, when I switch to creative writing tasks, it falls short of the 03‑25 exp model I’ve been using for ages. The tool feels like it’s tuned for developers, leaving my storytelling attempts feeling flat and uninspired. This mix of brilliance and disappointment made the overall experience just average.

I tried using Gemini 3 to create a coloring page, but it completely botched the task. When I asked it to review its own work and point out the errors, it just replied with another image—nothing helpful. The whole interaction felt odd and disappointing, leaving me frustrated that the model couldn’t even critique its mistake.