Nano Banana: Gemini 2.5 Flash Image Features & Launch

A codename that outlasted the launch

Google product manager Naina Raisinghani picked "Nano Banana" at 2:30 a.m. when a colleague messaged her that the team needed to submit a codename to LMArena, the public leaderboard where new models get tested anonymously before a vendor attaches its name to the result. She combined two of her own nicknames, Naina Banana and Nano, and submitted it as a placeholder. [4]

The model tested under that name for weeks in early August 2025 before Google confirmed it as Gemini 2.5 Flash Image on August 26. By Google's own later account, the placeholder had already taken over the conversation: social media discussion of the name outran the technical one, and Google leaned into it with banana-themed posts on X rather than fight the nickname. [4]

LMArena's own recap of the testing period reported over 5 million community votes across its Arena during that window, 2.5 million of them cast for this model alone, and the largest Elo score lead in the leaderboard's history: 171 points over the next-best system, with the model ranked first on both the Image Edit and Text-to-Image leaderboards. [3]

What shipped on August 26

Google's announcement described Gemini 2.5 Flash Image as an upgrade to the native image generation it had shipped in Gemini 2.0 Flash, aimed directly at complaints about quality and creative control. The stated capabilities: maintaining a character's appearance across multiple prompts and edits, blending several images into one composition, targeted local edits such as background changes or pose alteration from a plain-language instruction, and using the model's world knowledge to interpret a request semantically rather than as a literal pixel operation. [1]

It launched into the Gemini app for everyone, plus the Gemini API, Google AI Studio and Vertex AI, priced at $30 per million output tokens with each image counted as 1,290 output tokens, roughly 3.9 cents per generated or edited image. Google also made it available through the third-party platforms OpenRouter and fal.ai the same day, and every output carries an invisible SynthID watermark identifying it as AI-generated. [1]

Why consistency was the feature that spread

Atlas interpretation: Prior image-editing models tended to redraw more of a picture than a prompt asked for, so a face, a logo or a background detail would drift a little with every pass. A model that holds those details still across edits changes what editing means in practice: a user can issue a second, third and tenth instruction against the same subject instead of regenerating from scratch and hoping the result still looks like the same person or object. That is the mechanism behind LMArena's users nicknaming it a "Photoshop killer," and behind Google's own framing of it as multi-turn editing rather than one-shot generation. [3][1]

A prompt-engineering writeup published that November demonstrated the effect concretely: given a generated image and five distinct edit instructions issued together, such as swapping garnishes and adding background figures, the model applied all five while adjusting only the specific regions each one touched, including secondary effects like syrup pooling differently once a garnish moved. The author attributed the behavior to the model's autoregressive architecture, which he argued has an easier time targeting the specific tokens that correspond to one area of an image than earlier diffusion-based editors did. [2]