Gemini felt dumb on January 15, 2026.

I tried using Gemini’s Antigravity feature and it just crashed—every request ended with “agent execution terminated due to error.” It wasn’t just me; dozens of users were stuck, unable to get any work done. Google even acknowledged the bug in their forum, yet the status page stayed unchanged, which felt hugely frustrating and left my projects on hold.

I tried Gemini 3 Pro’s voice‑to‑text, and the first chunk transcribed flawlessly. But when I pause the recording to let background noise die down and then hit resume, the tool refuses to pick up where it left off. I end up copying the text, clearing the box, recording the next segment, and pasting everything together. It feels like a clunky workaround and makes the experience frustrating.

I fed Gemini 3 Flash and Pro over 60k tokens, hiding a “needle” phrase to test memory. Flash stopped following instructions after ~23k tokens, Pro after ~51.5k, but both later recalled the hidden cue and answered correctly. Their accuracy dropped as the context grew, with Pro ending around 72% at 64k tokens. The test showed the advertised 32k limit isn’t real, and context management still needs work.

it is a bit iffy... but has moments of utility that are insightful.

I tried using Google’s AI mode to get a quick fact, and it blurted out “450 years” instead of the correct “45 years.” The mismatch was glaring, especially since a regular search gave the right number. It felt oddly unreliable, and I’m left wondering how such a basic error slipped through and where I can report it.

I tried to craft a “night/noon” vibe story with the Nano Banana Pro, using each previous image as a reference. After a few generations, the pictures started turning noisy and developed an odd green tint. I’m stuck wondering why the model degrades like that and how I can stop it.

I tried to have a deep scientific discussion with Gemini 3.0 Pro, but the model kept insisting it wasn’t wrong, even when I cited reputable journal papers. Its stubbornness felt like it was rejecting solid evidence on purpose. After about twenty back‑and‑forth messages, it finally admitted a fundamental, systematic error that had been present from the start, leaving me frustrated and doubtful of its reliability.