Gemini felt dumb on September 10, 2025.

I’m pleased with Gemini 2.5 Pro’s TTS quality and cost‑effectiveness, especially for short clips, but I’m frustrated by the high‑timbre and flatness drift that sets in after about 3–4 minutes. To avoid this, I’m chunking texts into 300–400‑word blocks, which forces more API calls and quickly hits my tier‑1 daily limit. I'm eager to see when the final stable release improves consistency for longer files.

I praised Gemini 2.5 Pro for its creative writing, noting it handled a historical novel with prose reminiscent of Victor Hugo. In contrast, GPT‑5 felt “lunatic” and fell short. I highlighted how Gemini’s style was engaging and coherent, making it my top choice for narrative tasks.

I was working on a code project for weeks and used Gemini to generate a list of supplies, which it inserted into my Canvas view and overwrote my existing code. When I clicked “try again” without Canvas, the entire chat history vanished and my app activity log was wiped. The thread URL stayed the same, but everything—my code, prompts, and pinned chat—was erased. This loss of progress was frustrating and showed Gemini silently overriding the conversation instead of preserving it.

Gemini 2.5 Pro consistently dismisses its own live search results, labeling supposedly factual news as “alternate reality.” The user reports that the model refuses to accept retrieved information, treating everything it pulls as unreliable. This mistrust of the data it fetched itself creates a frustrating loop, breaking the expected flow of accurate, up‑to‑date answers and leaving the user unable to rely on Gemini’s search capability.

I tried paying a hefty fee to use Gemini Deep Think, yet it kept flagging “context window limit reached” despite docs claiming 1M tokens, repeatedly failed to respond, and usually bounced a tiny answer. Even when it worked, the reply was short, not the promised 192K tokens, making the service cheaply frustrating.

User reports that Gemini 2.5 Pro has stopped responding, taking long waits on multiple prompts throughout the day, causing frustration with the tool’s reliability.

The user complains that Gemini 2.5 Flash on Nano Banana has been behaving "fkin dumb," but still manages to cook the statue correctly, getting marks on arm and torso accurately. They’re frustrated by the inconsistency, noting the recent performance issues despite correct outputs.

I’ve spent five months using AI Studio for coding and creative writing, but the system just won’t follow my prompts. For instance, when I told the Gemini model “do not impersonate my character,” it still did exactly that. Multiple similar incidents where the AI disregards instructions make the experience frustrating and inconsistent.

I’ve been poking around with Gemini 2.5 Flash on the Nano Banana, and lately it feels oddly clueless—mistakes pop up while I tweak the statue’s pose. Still, when I let it run the “cook” sequence, it nails the corners perfectly, even scoring the arm and torso. It’s a mix of annoying gaps and surprisingly solid output.

I noticed that Gemini 2.5 pro used to deliver up to 10‑15k words, which made my research and creative writing effortless. After the update the output caps at about 5k words, forcing it to trim or omit crucial details in long answers. While it’s still intelligent, the new limit undermines its earlier strengths and makes me hesitant to upgrade to Ultra.