Nano Banana 2: Features, API Pricing and Text-Rendering Test

A Flash model wearing the Pro line's features

Nano Banana 2 is the product name for Gemini 3.1 Flash Image, an image generation and editing model Google released on February 26, 2026 across the Gemini app, Google Search, AI Studio, the Gemini API, Vertex AI, Flow and Google Ads. Its predecessor, Nano Banana Pro (Gemini 3 Pro Image, released November 20, 2025), stayed reserved for Google AI Pro and Ultra subscribers, so Nano Banana 2 is the version most people generating images in the Gemini app actually got. [1]

Google's own description groups the model's features under three headings: world knowledge (drawing on Gemini's knowledge base plus real time web search results to render specific real subjects), text rendering (legible text for things like marketing mockups and greeting cards, including translated or localized text), and subject consistency (keeping up to five people recognizable and up to fourteen objects intact across a single editing session). Output tops out at 4K. [1]

Atlas interpretation: The pitch is not a new capability so much as a new price point on an existing one. Every feature Google lists for Nano Banana 2 already existed in Nano Banana Pro; what changed is that a Flash-tier model, priced and paced for high-volume use, now carries them, and Google's own API pricing puts that model at roughly half the per-image cost of the Pro version. That is the kind of move that matters for a loop that calls an image model repeatedly, where a slower, pricier Pro call is a bottleneck rather than a one-off request. [1][2]

The text claim held up on signs, not on newsprint

PCWorld tested the text rendering claim directly on launch day, generating images meant to include billboards, headlines and newspaper-style layouts. The model rendered short, prominent text such as signage and headlines correctly, but text fidelity degraded when the request called for dense, small text, such as the body of a simulated newspaper article. [2]

Atlas interpretation: Text rendering in image models has historically failed hardest at exactly this scale: a few words on a sign are a much smaller prediction problem than a paragraph of body copy, where every character has to be right for the result to read as real. An independent test that separates the two, rather than taking the announcement's short demo examples at face value, is what turns a marketing claim into a bounded one: Nano Banana 2 improved on legible short text, and dense text is still where it breaks. [2]

One release in a monthly cadence

Nano Banana 2 arrived a week after Google released Gemini 3.1 Pro, the text and reasoning model the 3.1 generation is built around, and shares its version number with it. [1]

Atlas interpretation: The gap between the two releases is small enough that Nano Banana 2 reads as the image counterpart to a text model Google had already shipped that month, rather than a standalone event. Both belong to a pattern that had become routine by early 2026: a numbered model line generating a matching image variant within days, priced to sit inside an automated pipeline rather than a single design task. [1]