Google put a cheaper rung on its image-generation ladder on Tuesday, June 30, 2026. Nano Banana 2 Lite is the fastest and least expensive model in the Nano Banana family, tuned for developers who need to churn out large volumes of images without paying flagship rates per picture. It arrived alongside Gemini Omni Flash, a video model built to animate those same images, and together they mark Google's push to make generative media cheap enough to run inside high-volume product pipelines rather than reserve it for one-off creative work.
- Nano Banana 2 Lite generates 1K-resolution images in about four seconds at $0.034 each, roughly half the price of Nano Banana 2 and a quarter of Nano Banana Pro.
- Its formal model ID is gemini-3.1-flash-lite-image, and it replaces the original Nano Banana that ran on Gemini 2.5.
- Gemini Omni Flash launched at the same time for video, priced at $0.10 per second, and Google suggests chaining the two models.
- Text rendering inside images is better but still imperfect, as one early tester found the model misspelling the same phrase two different ways in a single picture.
What Nano Banana 2 Lite is built to do
Google's framing for Nano Banana 2 Lite is blunt. The company calls it the fastest and cheapest Gemini image model, engineered for velocity and scale, and everything about the release supports that positioning. The model produces text-to-image outputs in roughly four seconds with much lower latency than the heavier options in the lineup, which makes it suited to workflows where a developer or a user wants to spin through many iterations quickly instead of waiting on a single polished render.
The intended jobs read like a list of the places image generation gets expensive at scale: rapid ideation, interactive prototyping, visual drafting for later refinement, and budget-conscious applications that call the model thousands of times. This is the tier where per-image cost decides whether a feature ships at all. A social app that wants to generate a custom image for every post, or an e-commerce tool that drafts product visuals on demand, cannot absorb premium per-image pricing across millions of calls. Nano Banana 2 Lite is the answer to that specific constraint.
Despite the speed focus, Google insists the model does not collapse on quality. It says Nano Banana 2 Lite still delivers reliable prompt following, consistent character rendering across images, and readable text inside the generated picture. Those three properties are exactly the ones that fast, cheap image models have historically struggled with, so the claim is the interesting part of the pitch. Whether it holds up is a question the early testing starts to answer, and the answer is mixed.
The price ladder inside the Nano Banana family
The clearest way to understand Nano Banana 2 Lite is to look at where it sits in Google's pricing. According to The Decoder's breakdown, the family now spans three tiers by cost. Nano Banana 2 Lite runs $0.034 per 1K-resolution image. Nano Banana 2, the balanced all-rounder, costs $0.067 per image. Nano Banana Pro, aimed at complex professional work, sits at $0.134 per image. Each step up roughly doubles the price, so the Lite tier is about half the cost of the standard model and a quarter of the Pro option.
That structure tells you how Google wants developers to reason about the choice. If a task is high volume and forgiving of small imperfections, reach for Lite. If it needs the most faithful rendering of complex scenes or fine text, pay for Pro. The middle model covers the general case. It is a familiar shape for anyone who has picked between a flash and a pro tier on a language model, and Google is deliberately importing that mental model into image generation, where the cost gaps between tiers are large enough to change what a product can afford to do.
Nano Banana 2 Lite also replaces something. The original Nano Banana ran on Gemini 2.5 under the model ID gemini-2.5-flash-image, and the new Lite model, formally gemini-3.1-flash-lite-image, takes its place as the entry point. Simon Willison, who tracks these releases closely, noted in his hands-on writeup that Google describes the model plainly as the fastest and cheapest Gemini image model, engineered for velocity and scale. The version bump from 2.5 to 3.1 signals that this is not just a price cut on old weights but a newer generation offered at the budget tier.
Where you can use it
Availability is broad from day one. On the developer side, Nano Banana 2 Lite is reachable through Google AI Studio, the Gemini API, and the Gemini Enterprise Agent Platform. That covers the path from a quick prototype in the browser to a production integration wired into an enterprise agent stack. Google also confirmed the model is rolling into its own consumer surfaces, including AI Mode in Search, the Gemini app, NotebookLM, Google Photos, and creative tools such as Stitch, Google Flow, and Google Ads.
The spread across Google's own products is the tell. When a company puts a new cheap model into the image path of Search and Photos, and into its ads product, it is planning to serve image generation at a volume that only makes sense with a very low per-image cost. The consumer rollout and the developer pricing are two views of the same decision: Nano Banana 2 Lite exists because generative images are moving from a novelty feature into an always-on default, and always-on needs cheap.
Every image the model produces carries a SynthID watermark, Google's invisible marker for AI-generated content. That applies across the Nano Banana family and the new video model, so downstream tools can detect that a picture or clip came from a Google generator even after it has been edited or recompressed.
Gemini Omni Flash adds video to the mix
The second half of the announcement is Gemini Omni Flash, a video model that shipped to wider release the same day after an earlier preview at Google I/O. Its model ID is gemini-omni-flash-preview, and it turns text and images, along with short video references, into video output. Pricing lands at $0.10 per second of generated video, and the model currently produces or edits clips up to 10 seconds long, with Google saying longer durations are coming.
The more interesting capability is conversational editing. Rather than regenerating a whole clip from a new prompt, a user can steer Omni Flash through natural language, asking it to change elements of an existing video while text and graphics stay synced to the action on screen. As TechCrunch reported, Google frames the pair of models as tools for creative iteration and building end-to-end multimedia experiences, and it explicitly suggests chaining them: generate a still with Nano Banana 2 Lite, then animate that still with Omni Flash.
The limits are worth stating plainly, because Google states them too. Omni Flash does not yet support audio references. It will accept a video reference up to three seconds, but the company acknowledges the model does not process that reference with much accuracy. Character consistency across scene changes or camera movements is still a weak point, so a person or object may drift in appearance when the shot changes. This is preview-stage video generation, useful for short, single-shot clips and shaky on the longer, multi-shot narratives that professional work demands.
How the models work together
Google is not just shipping two models side by side, it is wiring them into a shared editing loop. Both Nano Banana 2 Lite and Gemini Omni Flash plug into an Interactions API that supports up to three sequential edits while keeping session context, so a user can refine an image or a video across a few turns without starting over each time. That continuity is what turns a one-shot generator into something closer to a design tool, where each request builds on the last.
To show the idea off, Google published demo applications that stitch the models into concrete tasks. One called Anywhere transforms the location in an image. Another named Space Lift handles interior design edits. The third, Omni Product Studio, converts static product images into what Google calls cinematic e-commerce videos, which is the clearest commercial pitch in the set. An online store could feed a flat product photo into the pipeline and get back a short animated clip for a listing or an ad, generated at a few cents rather than a filmmaker's day rate.
Does the quality hold up?
Cheap and fast only matters if the output is usable, so the early hands-on impressions carry weight. Willison ran the same prompt he had used on earlier Nano Banana versions back in April, a scene involving a raccoon holding a ham radio, and reported that the new model's output was better than what those older versions produced. His broader read was that Nano Banana 2 Lite is a genuine quality step up at the budget tier rather than a stripped-down fallback.
The rough edge showed up in text. Willison generated a detailed Where's Waldo-style woodland festival illustration, full of animal characters and busy environmental detail, and the composition held together well. But the model rendered the phrase Forest Festival incorrectly in two different ways within the same image. That is a familiar failure mode for image generators, and it is notable here precisely because readable text was one of Google's three headline quality claims. The model is closer to getting text right than its predecessors, and it still cannot be trusted to spell a two-word sign correctly across a single busy scene.
The practical takeaway is to match the model to the job. For thumbnails, backgrounds, mood boards, and drafts that a human or a Pro-tier model will finish, Nano Banana 2 Lite looks like a strong default at its price. For anything where a legible headline, a precise logo, or exact on-image text is part of the deliverable, the budget tier is not the place to stop. That is not a knock on the model so much as an accurate reading of where a four-second, three-cent image lands on the quality curve.
The conversational editing loop in practice
The Interactions API is where the two models stop being isolated generators and start behaving like an editor that remembers what you asked for. Google caps a session at three sequential edits with retained context, which is a modest window but enough to cover the common refine-and-adjust rhythm of real work. A user generates a base image, asks for a change, then asks for another, and each step sees the prior state rather than treating every prompt as a fresh start. That memory is the difference between fighting a generator and directing one.
The demo applications make the loop concrete. In Space Lift, the interior-design demo, a person can take a photo of a room and iterate on furnishings or finishes across a few turns, watching the space change while the underlying layout stays put. In Anywhere, the same mechanism swaps the location behind a subject. Omni Product Studio carries the pattern into video, taking a static product shot and producing a short animated clip suitable for a storefront. The common thread is that a non-expert can reach a usable result through plain requests, without touching a layer panel or a timeline, and at a cost measured in cents rather than the hours a designer would bill.
The honest caveat is that the three-edit ceiling and the preview-stage video limits keep this short of a full creative suite. You cannot run a long chain of refinements, and Omni Flash's weakness at holding a character steady across cuts means multi-shot video still needs human cleanup. The loop is genuinely useful for the first draft and the quick adjustment, and it is not yet a replacement for a designer on anything that has to be exactly right.
What the Gemini 3.1 lineage signals
The model ID gemini-3.1-flash-lite-image places Nano Banana 2 Lite firmly in Google's current generation, not a rebadge of older weights. The original Nano Banana rode on Gemini 2.5 as gemini-2.5-flash-image, so the jump to 3.1 tracks Google's broader model cadence and suggests the image tiers now advance in step with the language models that share the Gemini brand. That alignment matters for developers because it hints the image and text capabilities will move together on a predictable schedule rather than drift on separate tracks.
It also clarifies Google's competitive posture. By fielding a fast budget image model, a balanced standard one, and a premium professional tier, all watermarked with SynthID and all reachable through the same API surface, Google is selling a full ladder rather than a single hero model. The bet is that most developers want to pick a price-quality point per task and switch freely between them inside one integration. Nano Banana 2 Lite is the floor of that ladder, and its arrival at $0.034 an image is Google planting a stake at the low end of a market where cost per generation increasingly decides what gets built.
Why the low price is the real story
The headline number here is not the four-second latency, it is the $0.034 per image. At that cost, image generation stops being a feature you meter carefully and starts being something you can run on every request. A product that shows a user ten variations instead of one, or regenerates a visual every time a caption changes, becomes economically sane when each image is a fraction of a cent. The same logic that made cheap language-model tiers reshape which apps could afford AI text is now arriving for images.
Google's own product rollout is the proof of intent. Putting Nano Banana 2 Lite into Search, Photos, and Ads only pencils out if the per-image cost is low enough to serve at that scale, and $0.034 is that number. Meanwhile Gemini Omni Flash at $0.10 per second points the same trend at video, though the preview-stage limits mean video is a year or two behind images on the reliability curve. The direction is unmistakable: generative media is sliding down the cost curve fast enough that the interesting question is no longer whether a product can afford to generate an image, but what it does once generating one is nearly free.
For developers weighing the options, the decision comes down to volume and tolerance. High-volume, quality-forgiving work now has a credible budget option that does not embarrass itself on prompt following or character consistency. Precision text work and complex professional renders still belong on the pricier tiers. And anyone building a media pipeline can now generate a still and animate it through one connected API, at a combined cost that would have looked impossible a year ago. That combination, cheap images and cheap short video from the same vendor through the same interface, is what makes this launch more than a routine price cut.