Its Own Exec Had Already Said:
Nano Banana’s Launch Was a ‘Success Disaster’
Give your enemy the cheap thing from day one and he never complains about it. Give him the good one for three months and then take it back, and now you have turned a paying customer into an enemy.
Google worked through that lesson this year with an image model, and the argument it started has been running ever since, with neither side landing a punch.
What Google Changed On 26 February
Nano Banana Pro was released on 20 November 2025 as the Gemini 3 Pro Image, drawing or editing whatever you asked the Gemini app for during the first three months post-launch.
This was followed by Google releasing Nano Banana 2 (Gemini 3.1 Flash Image) on 26 February. The 3.1 version is set to Flash instead of Pro by default. Google explained that the Flash version is faster and cheaper for them, and that it also follows instructions better and has sharper detail.
Pro was never taken away, it was instead moved to a cupboard.
According to Google’s help page, you generate with Nano Banana 2, and paid subscribers regenerate with Nano Banana Pro. In the app, this works out to a three dot menu and a Redo with Pro button. So you don’t pick the model at the start. You ask the cheap one first and then appeal the result using the expensive one.
Google sent nothing to the people paying nineteen ninety nine a month for it. They found out by noticing their pictures had changed.
Josh Woodward Was Already Calling The First Launch A “Success Disaster”
This is the thread I had missed until I went back through last year’s coverage.
Josh Woodward, who now runs Gemini at Google, called the launch of the original Nano Banana in August 2025 a “success disaster“ in an interview for the Wall Street Journal, which reads like a bad joke until you see what happened to them.
Thirteen million people opened the Gemini app for the first time in the four days after launch that September, and more than five billion images had been generated by mid October. Woodward explained that Google could not find enough computing power to keep the thing running.
They ran out of chips.
Put that next to the February swap, where Google swapped a Pro model for a Flash model, and you see Flash exists because it costs less to serve at scale. Five weeks before the swap, on 21 January, the Pro API fell over for five and a half hours, with 4K requests timing out while 1K and 2K requests continued, which Google attributed to its TPU compute provisioning.
Google has never linked the swap to capacity, but the head of their app has said capacity was their issue, and the swap moves in the direction that fixes it.
What Paying Subscribers Reported Afterwards
If you read through the comments people have been posting on Reddit, on X, in the Gemini Apps community and in Google’s AI Developers forum, it’s the same basic problem, reworded, time after time, for months. It’s one topic.
Faces getting ruined.
“The image consistency is trash. They look about 50 years older than they are, the skin looks completely flat and fake, and annoyingly, they keep getting spotty blemishes that weren’t there to begin with.”
I agree, Nano banana pro is unable to replicates faces as well as the base version. Very disappointed. The pro version tends to embellish and make faces more conventionally attractive.
The recent updates to Nano Banana 2 have severely ruined the image quality and realism. The model is now producing outputs with cheap, plastic textures. It went from generating hyper-realistic images that anyone could mistake for real photos to producing absurd, low-quality, cartoon-like characters.
Please understand that just because something is fast doesn’t mean it’s good. Speed shouldn’t destroy quality. We need the previous realism and depth back. Please fix this regression or give us an option to choose quality over generation speed.
Looks like today Gemini and Nano Banana been creating really bad images even using your same prompts, it looks like they did a reset or something, but looks like the model is starting learning again like it was when they launched it. Everytime I have to tell Gemini to try again because the first image come up buggy or with no sense. Never happened before the previous days or months. All this happening on top of the Safety Filters refusing to create anything they think is explicit content as well.
Someone is messing with the buttons in Google because I don’t remember seen Nano Banana working this bad in long time ago.
Skin going flat, people ageing a decade in the edit, and marks turning up that never existed in the photograph they uploaded.
Then there is the escape hatch, which drew its own complaints. What good is a redo button when the thing hands back nothing a good share of the time, one subscriber wrote.
Google hasn’t claimed that it changed the model, and from what I can see, it didn’t change. Pro still does what Pro did. It just stopped being the thing you land on.
Nano Banana 2 Ranks Eighth, Pro Ranks Eleventh
The board that I published a few days ago makes the complaint threads that much harder to defend.
Nano Banana 2 is ranked eighth, at 1101 Elo on slightly less than fourteen thousand votes, at a cost of $67 per thousand images. Nano Banana Pro is ranked eleventh, at 1094 Elo on slightly under thirteen thousand votes, at a cost of $134.
Google replaced it with a higher ranked model, judged on a similar pile of votes, for half the cost.
February gave neither side anything to argue with. People describing skin texture met people saying it looks fine to me, and both groups carried exactly the same weight of evidence, which came to nothing. Seven months of head to head voting settles it.
By the numbers Google made a call it can defend. Cheaper, faster, ranked higher, and it claimed the recommended replacement slot when Google shut down the previous Imagen generation models on 17 August.
The Board Scores Edits, The Complaints Were Portraits
The board scores edit work: replacing a background, editing out an object, re-lighting a frame, dropping a product shot onto white, pulling a bin out of a street shot. Voters compare that sort of thing, and a Flash model built around following instructions does it well.
The complaints came from portraits. Nearly every one of them.
Two different jobs, and the gap between them is just where a Flash model gives ground. Skin texture, all the little bumps that make a face read as a face, the restraint to leave alone something that already worked. Reasoning does that, which is why the Pro model still beats it there while losing out on the board.
Both sides described their own work accurately, and neither said what kind of work it was.
I looked for multiple months for someone who separated the two, and it is mostly not there.
Redo With Pro Runs On The Free Model’s Quota
Miss this one and you are out of money.
Redos are limited per day on Google’s plans, 50 for AI Plus, 100 for AI Pro and 1,000 for Ultra, and the Basic plan has no redo lane at all.
The catch is underneath. Because you are limited by the free model’s quota, the day you finish your Nano Banana 2 generations will kill your Pro redo too, since a redo runs as a second pass over a generation you already spent.
Your free allowance rations the thing you pay for.
Adobe Confirmed The Credit Errors In February
February hit more than Google’s own app.
Creative Cloud subscribers running Nano Banana Pro through Firefly and Photoshop started hitting the “Something went wrong” error, and Adobe took the credits for the failed generation anyway. One user said:
“makes the feature practically unusable (and quite unfair for the user)”
The thread ran to 55 replies. Two Adobe community managers confirmed publicly that some Creative Cloud users were seeing high error rates and delays.
Which Model Fits Which Job?
You edit products for a living. Google did you a favour you never noticed. You moved from eleventh to eighth, and your bill halved.
Edit people and February cost you something the board will never register, because the board asks a different question than the one you care about.
Want Pro as a model instead of as an appeal against the cheap model’s output, and the Gemini app has stopped serving that. Front ends like Nano Banana Pro still let you pick it outright, which walks past the buried menu and the quota gate together. I count on that sort of thing because pushing one photo through several models from a single login is preferable to juggling three accounts, and it handles this as a side effect.
What bothers me is that the answer had already been there for months. The votes were stacking up all that time.
I publish the board weekly, but this is the first time I have ever seen it read as the answer to anything.
