Testing Catalog notes that Nano Banana 2 model “exhibits a new multi-step generation workflow: it spends considerable time planning the output, then generates an image, reviews it via built-in image analysis, identifies and corrects any errors, and iterates as needed before delivering the final result to the user.” This hasn’t been seen before, and can improve the accuracy of images. The model was available via Media IO in preview for a few days, identifying itself as a Google-made model. According to the same Testing Catalog, the model was pulled from Media IO on Sunday.
Update: Nano Banana 2 option was removed
— TestingCatalog News (@testingcatalog) November 9, 2025
Other users on X noted that the Nano Banana 2’s test seemed to lack some safety guardrails. As you’ll see in the various samples below, testers were able to use celebrities and political figures to create stunning images. For example, one photo shows the purported results of the 2027 presidential elections as presented on CNN, with Trump projected to win his third term in office. Other Nano Banana 2 examples show Google’s AI model creating Windows screenshots that look just like Microsoft’s operating system, complete with text generation. Here are some Nano Banana 2 samples you can take a closer look at:
Nano Banana 2 is AVAILABLE on https://t.co/KB1857eejK, here is an output for proof: pic.twitter.com/M3JfaHKR1T
— Leeham (@Liam06972452) November 8, 2025
Nano Banana 2 completely smashed both the clock AND full wine glass tests in ONE IMAGE.
“11:15 on the clock and a wine glass filled to the top”
Holy. Shit. https://t.co/SedbYL92PX pic.twitter.com/q47HW04ZsH
— leo (@synthwavedd) November 8, 2025
Generated by Nano Banana/GemPix 2. No reference image, just txt2img.
Holy smokes. https://t.co/SedbYL9AFv pic.twitter.com/NwROjGyJsd
— leo (@synthwavedd) November 8, 2025
this one has even fewer issues. insanity
“Generate a screenshot of a windows 11 desktop, with google chrome open, showing a YouTube thumbnail of Mr. Beast on https://t.co/1VFaM2aqoa” https://t.co/TE1H0PI9HS pic.twitter.com/IoGopKco8b
— leo (@synthwavedd) November 8, 2025
Here’s another Nano Banana 2 image from yesterday before it stopped blocking political prompts that makes me think this model may be reasoning before generation.
Notable here is the correct percentages and raw totals including the “difference” value, and the correct headshot of… pic.twitter.com/j8MZdmF33M
— leo (@synthwavedd) November 9, 2025
Nano Banana 2 is absolutely insane.
I’ll share more outputs soon, hopefully they don’t censor the model before release. This is crazy
We’re about to get Gemini 3.0 Pro, GPT-5.1, GPT-5.1 Pro, Nano Banana 2, and Genie 3, which I’m really excited about.
Credit to @URUBONZ_. pic.twitter.com/F0BblzoGep
— can (@marmaduke091) November 9, 2025
An absolute fucking gem from @fleebdoo, again generated by Nano Banana 2 LMAO
The model will not be this uncensored at launch unfortunately https://t.co/SedbYL92PX pic.twitter.com/OQzzfjGxy5
— leo (@synthwavedd) November 8, 2025
Wild times.
Nano Banana 2 apparently leaked for a few hours today on some platforms, and the limited results by a small handful of users have been pretty insane.
The model was apparently from an uncensored, slightly older checkpoint.
When it actually launches it won’t be this… pic.twitter.com/Xe5Guddg9i
— Roberto Nickson (@rpnickson) November 9, 2025

