Gemini 3: initial usage thoughts

Gemini 3 feels tighter, more consistent in long responses, and more willing to pull in relevant images and references than earlier versions.

Published: Wednesday, November 19th 2025

1 min (222 words)

ajfisher - Gemini / Nanobanana

I have no idea what Google did in the jump from Gemini 2.5 to 3, but responses are significantly tighter and the tone is way more consistent across a long, mulit-turn sessions.

One of my observations with Gemini in the past is that a single response could often feel like 2 or 3 different conversations jammed into one, and it definitely jumped around between responses quite a bit too. Using it often felt like it was trying to impress you with how much it could do, but didn’t care much about how the five things it was recommending were synthesised back in a way that was cohesive.

It now seems far more likely to drop in contextually relevant non-generated images with references (no doubt leveraging their huge search assets) which is a welcome addition to simply dumping link refs. Even adding links feels more consistently applied and useful.

Apparently, it still can’t make images that aren’t square - which is deeply annoying. Though as an improvement you can get letterboxed images now rather than it simply cropping off the content like it did previously.

An alternative hack is to upload a blank image (yes literally just a white or transparent rectangle) that is the aspect ratio you’re after and it will then “edit” your source image with the generated content. Here’s hoping they fix that soon.