Google shipped Nano Banana 2 Lite on 30 June 2026 alongside Gemini Omni Flash, and the pitch is refreshingly narrow: the fastest and cheapest image model in the Gemini family. Roughly four seconds a picture, $0.034 per 1K-resolution image through the API, free if you just want to poke at it in the Gemini app or AI Mode in Search. It is the trimmed-down follow-up to Nano Banana 2 (Gemini 3.1 Flash Image), which arrived in February as Google's default image generator across its consumer products.
Speed-and-price claims are easy to make, so on 19 July I called gemini-3.1-flash-lite-image directly through the Gemini API and gave it the brief I use to embarrass image models.
The test: make it write
Text is where image generators still humiliate themselves. So the prompt asked for a flat-design poster on a white background, vermilion accent, carrying two exact lines: "AI JUST DROPPED" and "tracked the day it drops". No aspect ratio, no other constraints — I wanted to see what it would do unsupervised.
It came back in 4.3 seconds. Both lines rendered perfectly: correct letterforms, correct words, nothing garbled, nothing melted into pseudo-Latin. That is the single most useful thing I can tell you about this model. The flat-design direction and the vermilion accent were both followed accurately, from one prompt, with no retries.
The API's own accounting for that run: 46 prompt tokens in, 1,497 candidate tokens out, of which 1,120 were the IMAGE modality.
Where the "Lite" shows
Not in quality — in control. Because I never specified a ratio, it picked one, and picked tall portrait. That is a perfectly reasonable default and completely wrong if you were expecting a 16:9 social card. Say the shape out loud or accept a surprise.
The file that came back was image/jpeg, 563 KB, with C2PA/JUMB provenance metadata embedded. JPEG is a defensible default for photographic output and a slightly annoying one for flat vector-ish graphics, where you would rather have crisp PNG edges and an alpha channel. The provenance metadata is a good thing — it just means the file is quietly labelled as AI-generated wherever it travels, which is worth knowing before you drop it into a client deliverable.
Who it's for
Anyone generating images by the hundred where legibility matters more than art direction: social cards, blog thumbnails, deck backgrounds, quick mockups, placeholder assets in a build pipeline. At three-and-a-bit cents an image with four-second turnaround and text that actually reads, the economics stop being a consideration at all — you iterate by regenerating rather than by prompt-engineering.
Who should skip it
If you need transparency, lossless output, or fine-grained compositional control, this is the wrong tier — step up to the full Nano Banana 2, or to a model you can steer with more than prose. If you need weights on your own hardware, this is the wrong model entirely: it is closed, hosted, and Google-only.
One caveat on my number. This score comes from a single, deliberately hard generation — a real hands-on run, not a spec-sheet read, but one prompt is one prompt. It tells you the model is fast and can spell. It does not tell you how it handles faces, hands, complex scenes, or a hundred consecutive briefs. I will revisit as I put more through it.