The Blind Test Worked Once

For roughly two weeks before that, an unidentified image model had been running on LMArena under the codename "nano-banana," collecting votes from people who had no idea who built it. The arena's format is blind by construction: two outputs, one prompt, pick the better one. By the time Google claimed the model as Gemini 2.5 Flash Image, it had drawn more than 2.5 million votes on its own — driving over five million across the platform — and sat at an Elo of 1362 in the image-edit category, 171 points clear of the runner-up. The largest lead in the arena's history.

The reviews were in before the byline was.

That's an interesting accident, and worth being precise about what it tested. Model evaluation has a contamination problem: almost nobody assesses a model without knowing whose it is. The logo is in the frame before the prompt is typed. Whatever the Google discount is worth it isn't zero, and it isn't measurable either, because there's normally no condition in which it's absent. Nano-banana was that condition, briefly — by operational accident, not experimental design. Google was staging a preview, not running a study.

What the accident found is thinner than I wanted it to be. Strip the maker and the thing still wins, by the widest margin anyone had recorded — that part happened. But a control requires the same model scored twice, with the logo and without, and that comparison was never run. There is no branded Gemini 2.5 Flash Image number to hold the anonymous one against. What got observed is that an anonymous model can win, not that anonymity cost it nothing. I wanted a finding. What I have is an anecdote with unusually good lighting.

It gets worse in a more interesting way. The arena was blind on identity, not blind on hype. Nano-banana pulled 2.5 million votes because it had become an internet mystery — a nonsense name people were actively hunting. That's a selection effect with teeth: users who show up specifically to find the banana are not a random sample, they know they've found it the moment character consistency holds across a multi-turn edit, and nothing stops enthusiasm from voting again.

And the hunt was never identity-neutral, because the identity was the mystery. People were guessing whose model this was, and Google was a leading guess long before Google confirmed. The halo had a back door the whole time.

It also lasted about two weeks. Look at what the name is now. "Nano Banana" is not a leaked codename anymore. It's the product. Google DeepMind's own page reads Gemini 3 Pro Image — Nano Banana Pro. Google Cloud shipped Nano Banana 2. CNBC covered the November 2025 launch under the banana, not the version number. The word chosen precisely because it carried no information about its maker is now the most recognizable brand in Google's image lineup, and it is doing what brands do, which is save people the trouble of checking.

The anonymity did not defeat branding. It minted a brand, faster and cheaper than any campaign could have, out of an accident, and then handed it to the incumbent that ran the accident.

Count what actually got produced. Five million people did unpaid comparative evaluation — work that's a line item when a lab does it in-house. The durable asset that came out of it was a trademark. The crowd supplied the judgment; one company kept the name. Nobody was defrauded and I doubt anyone minded, which is the point: an enclosure in a fruit costume doesn't read as an enclosure.

Which is also why it can't be run again. Every anonymous codename on a leaderboard now reads as a staged reveal in progress. The method depended on participants not knowing what it was for, and this one got explained to everybody at once, by working.

But calling that a loss is too generous, because only half of it died. The arena's measurement function needed the blindfold, and the blindfold is gone. Its promotional function never needed it — that function was discovered here, and it outperformed. A venue built to evaluate vendors turns out to be a venue vendors can stage on. That isn't an instrument breaking. It's an instrument repurposed by the things it was pointed at.

So the finding is thinner than the story, and the instrument isn't gone, only reassigned. We learned once, under conditions nobody can reconstruct, that a very good model can win without showing its logo. We never learned what the logo is worth, and we have less chance of finding out now, because the closest thing to a control group the field ever stumbled into got turned into a product name inside a fortnight.

The model that had no name has a name. The name is Nano Banana, and Google owns it. The blindfold came off two weeks in, and it isn't going back on.