Gemini felt dumb on February 20, 2026.

I tried the 3.1 Pro in Antigravity Free and it was painfully slow, so I upgraded. The pro version talks nicely and runs smoothly, but its implementation often falls short—I have to double‑check its output and make fixes. Opus feels collaborative, Codex is assertive and gets things done, yet Google’s Studio with Repomix still seems to produce cleaner code. The experience was a mix of helpful moments and frustrating re‑work.

I tried both Claude and Gemini 3.1 Pro side‑by‑side and found the Google suite surprisingly strong. Claude still works great for me at work, but Gemini handled my home‑lab scripts, side‑project tweaks, and even the new antigravity features without a hitch. The all‑in‑one Google One plan—storage, AI, Notebook LM, CLI—felt like a complete package, so I finally dropped one subscription. The experience was smooth and satisfying.

I tried generating text with the new half‑new model, expecting longer outputs like the 3.0 Pro gave a couple of days ago. Instead it kept cutting off early, and the length was noticeably shorter than before. The inconsistency was obvious since yesterday, making the tool feel unreliable and frustrating when I needed more extensive responses.

I kept running into syntax errors when using the 3.1 pro model and it was driving me nuts. After digging, I suspect the tokenizer got messed up during training, which explains why the code kept breaking. The constant need to debug the AI‑generated snippets was frustrating, and it made the whole experience feel clunky and unreliable.

I tried importing PDFs into Gemini 3.1, expecting it to read and search their contents. Instead, the model kept responding that it could only see file names and metadata, even when using notebooklm. The repeated “I don’t have the backend document‑reading tool” messages were frustrating, making the feature essentially unusable.

I was in the middle of building my app when Google AI Studio suddenly hit me with a “You are out of free prompts” notice, even though I have a paid API key and billing is active. After that, every build returned an internal error. I tried remixing, restarting, and checking the project settings, but the error kept looping, leaving me stuck and frustrated.

I tried using AI Studio’s 3.1 Pro to generate code, but the output kept getting truncated and malformed, breaking the syntax. It worked fine in the Gemini app, so the problem seems isolated to the Studio interface. The constant formatting glitches were irritating and made me waste time fixing the code manually.

I tried to build Roto in AI Studio and the experience was downright horrible. The tool kept crashing, gave nonsensical suggestions, and completely derailed my workflow. I felt frustrated and stuck, wasting time fixing errors that shouldn't have existed. Overall, the AI's performance was a massive disappointment.

I tested Gemini 3.1 Pro on the Vending‑Bench and was disappointed to see it lag behind Gemini 3 Pro. The responses were less accurate, the tool missed key details, and overall it felt like a step back. The experience left me frustrated because I expected an upgrade, not a decline.

I keep trying to give Gemini specific instructions, but it often just skips them and goes off on its own track. It’s frustrating because I have to redo prompts or clean up its output, feeling like the model isn’t respecting my guidance. The inconsistency makes the workflow slower and leaves me doubting its reliability.

I tried using 3.1 pro in AI Studio to tidy up a handful of ~100‑line Python scripts. Instead of catching hidden logic bugs, the model spat out code riddled with absurd syntax mistakes—sometimes it left the right side of an assignment completely empty. The output didn’t even run, and I had to fall back to 3.0 flash to get anything usable, which was frustrating and a waste of time.

I’ve been playing around with Gemini 3.1 and, honestly, it feels like a solid .1 update. It exceeded my expectations in small ways, even if it isn’t a massive .5/.0 leap. My own tests are pretty saturated now, and overall it feels like a step‑change, just not a game‑changer.

I asked Gemini 3.1 Pro to output a tiny snippet—exactly `locale: ['en', 'es'],`—and it didn’t obey. Instead of the precise line I needed, the model garbled or omitted it, forcing me to troubleshoot manually. The bug felt odd and irritating, especially since it wasn’t a UI glitch but a core generation flaw, leaving me doubting its reliability for simple, exact‑text tasks.

I kept waiting for the AI's research feature to finish, only to hit an error each time. The worst part was that I couldn’t just hit regenerate—I had to start a fresh prompt and run the whole research process over again, which was time‑consuming and irritating.

I tried using NanoBanana Pro to create an image for my research, but the service kept glitching and wouldn’t produce anything. Then I switched to Vertex and faced absurd wait times—sometimes up to five hundred seconds—or it just got stuck on “thinking.” The whole experience was exhausting, and I have no way to tell if there’s a wider outage.

I noticed Gemini 3.1 suddenly got the "car wash" question right, but it felt like they only patched that one example instead of fixing the underlying logic. I’m planning to throw more tests at it, hoping the fix isn’t just a band‑aid. The experience left me skeptical—briefly amused but still frustrated by the shallow fix.

I’ve been testing Gemini 3.1 Pro and noticed it hallucinates far less than the older Gemini 3 or even ChatGPT 5.2. When I ask for bibliography and in‑text citations, it actually follows the titles and drops proper links. It still slips up occasionally, but a strict prompt to “provide the link” dramatically reduces the errors. The improvement feels surprisingly reliable.

I was hoping Gemini 3.1 would finally fix the hallucination problem, but it was a huge disappointment. It still blithely throws out made‑up facts without any verification, just like the 3.0 version. The answers feel confident yet completely unreliable, so I can’t trust anything it says unless I already know the truth. This level of nonsense is frustrating and risky.

I tried Gemini 3.1 Pro with my paid API key and was blown away by the screenshots I posted. The benchmark hit 100% on the frontier test, and the SWE scores were off the charts. Seeing those results felt like a revelation—everything ran flawlessly and even exceeded what I expected from a commercial model. The experience was exhilarating.

I tried sending a text file of my creative writing to Gemini 3.1 for feedback, and for the first time I hit a rejection. The file was well within the context window and contained nothing unsafe—no NSFW, violence, or crime. When I switched to the older 3 Fast model it worked fine, and other Chinese models assessed it without issue. The unexpected block felt puzzling and frustrating, making me wonder if Google has over‑censored the new version for writers.

I’ve been relying on Gemini 3.0 Flash/Pro for months to spit out perfect JSON for my pipeline, and it never missed a beat—Pydantic validated every run. After upgrading to 3.1, my logs turned red: missing required fields, hallucinated keys, constant validation failures. I tried re‑phrasing prompts and tweaking system instructions, but the model just can’t stick to the schema. It feels like a regression that broke the core functionality I trusted.

I tested Gemini 3.1 Pro using the suggested prompt to describe a nonexistent image, expecting the hallucination bug to be gone. Instead, the model kept inventing detailed captions for random filenames, proving the issue remains. The experience was irritating—each try produced convincing yet false descriptions, showing the tool still can't handle missing visual data.

I tried to launch a new app with 3.1, expecting it to just code along. Instead, it paused me and reminded me of the website I’d promised myself to finish. That unexpected pushback felt like an actual advisor, not a mindless “yes man.” The tool’s behavior was surprisingly supportive, keeping me on track with my bigger goals.

I was blown away when Gemini 3.1 finally cracked the notorious seahorse problem—I’d been stuck on it for weeks. The moment it produced a correct solution, I felt a rush of excitement and disbelief, like a breakthrough in real AI. Seeing it handle such a complex task so effortlessly made me think, “This is actually AGI,” and left me thrilled and optimistic about what’s next.