Nano Banana 2 API

Nano Banana 2 is the community name for Google's Gemini 3.1 Flash Image model — API id gemini-3.1-flash-image-preview. It is the fast, high-throughput tier of Google's image family: text-to-image and reference-image editing at 1K, 2K and 4K, tuned for speed and unit cost rather than for maximum fidelity. On apimodels.app, Nano Banana 2 costs $0.05 per image at 1K and 2K, and $0.08 at 4K.

The short answer for anyone comparing tiers: if you are producing hundreds or thousands of images a week — product shots, social cards, video thumbnails, ad variants, localized banners — Nano Banana 2 is the tier that fits, because the cost of a mistake is five cents and a few seconds. If a single hero image has to be perfect on the first try, step up to Nano Banana Pro (Gemini 3 Pro Image) at $0.068/$0.13. Both models answer on the same endpoint with the same parameters, so switching between Nano Banana 2 and Nano Banana Pro is a one-string change.

What makes Nano Banana 2 interesting is not that it is affordable in isolation — it is that the quality gap to the premium tier narrowed far more than the price gap did. Gemini 3.1 Flash Image renders legible in-image text, follows multi-clause prompts, and accepts reference images for editing, which a year ago were premium-tier features. That is why Nano Banana 2 has become the default workhorse for batch image pipelines rather than a budget fallback.

Google bills Gemini 3.1 Flash Image by image tokens — $60 per million — which works out to $0.067 for a 1K image, $0.101 for 2K and $0.151 for 4K on the standard tier. We charge one flat $0.05 for both 1K and 2K, and $0.08 for 4K. That is where the headline number comes from, and it is worth stating precisely rather than rounding: Nano Banana 2 is 25% less than official at 1K, 50% less at 2K, and 47% less at 4K.

The 2K row is the one to look at. Google charges $0.101 for a 2K image; Nano Banana 2 here is $0.05, so 2K is where the gap is widest — you pay less than half. And you pay it in real time: post a request, get the image back in seconds. Anyone quoting a single blended discount for Nano Banana 2 is rounding away the part that actually matters, which is that the savings grow with resolution.

Two more things that change the real bill. Failed generations are never charged — you pay for images you actually receive, not for attempts. And there is no minimum, no monthly platform fee and no seat licence: an account that spends $3 this month pays the same per-image price for Nano Banana 2 as one spending $3,000.

One detail first, because it is the most common 400: the model string for Nano Banana 2 is gemini-3.1-flash-image-preview. The word nanobanana2 is this page's slug and our internal channel name, not an accepted model id. Send the Gemini id and you are calling Nano Banana 2.

Everything runs through one endpoint and one bearer key: POST /api/v1/images/generations. By default it creates a task and returns a taskId immediately, which you poll with GET /api/v1/images/generations?task_id=... or receive on a callback_url — the right shape for batch and queue work. For an interactive UI, add an OpenAI-style size (for example 1024x1024) and no callback_url: the same request then waits for the image and answers in the OpenAI shape, {created, data:[{url}]}, with response_format set to url if you want a link rather than base64. Editing uses that same endpoint — pass image_urls alongside the prompt. The body is model, prompt, optional image_urls, aspect_ratio (1:1, 4:5, 3:2, 16:9, 9:16, 2:3, 21:9 and others) and resolution set to 1k, 2k or 4k.

The Nano Banana 2 response hands back a hosted URL rather than base64 by default, so you can drop it straight into a CMS or a Slack message. Those result files live for 30 days — long enough for a pipeline to pick them up, not a permanent CDN. Copy anything you intend to keep into your own storage as part of the same job. If you want to see the exact request before writing code, the playground at the top of this page issues the real call with your key and shows the payload.

Nano Banana 2 is strong on short in-image text — a headline, a price tag, a label, a two-line caption — which is the capability that used to force people onto premium models. It follows compositional instructions (place the product left of frame, keep the background matte white, shoot slightly above eye level), it holds a consistent palette across a batch, and it accepts reference images so you can restyle or extend an existing asset instead of describing it from scratch.

Where it gives ground to the premium tier: dense paragraphs of small text, complex charts with many labels, precise brand-mark reproduction, and scenes where one specific object must stay pixel-consistent across a dozen variants. Those are Nano Banana Pro jobs. A pragmatic split that works well in production is to draft with Nano Banana 2 and finish the few images that ship at full size with Nano Banana Pro — the drafting stage is where the volume is, and that is where the $0.05 price does the work.

Worth knowing before you design around it: 4K is a real render at $0.08, not an upscale of a 1K image; every request is independent, so identical prompts do not return identical images unless you feed reference images; and content moderation is Google's, applied upstream, so a refused prompt is refused for reasons we do not control and do not bill you for.

The four jobs below are where Nano Banana 2 volume actually comes from on this platform. The reference images are pulled from our open prompt library, which is written for GPT Image 2 — every prompt in it runs on Nano Banana 2 unchanged by swapping the model string, which is the fastest way to see how the two models differ on your own subject matter.

E-commerce and product catalogues

A catalogue refresh means one product photographed against six backgrounds, in three aspect ratios, for four marketplaces. At $0.05 a frame, Nano Banana 2 turns that from a photo-shoot budget line into a script that runs overnight. Feed the existing pack shot through image_urls and the model keeps the product while replacing the surface, the light and the props.

Social posts and campaign variants

Social is a volume game: one message, twenty visual treatments, tested and discarded weekly. Nano Banana 2 handles the short headline inside the frame, so the copy and the art arrive together instead of going through a layout step. Twenty variants cost a dollar, which is affordable enough that testing beats deliberating.

Video thumbnails at channel scale

Thumbnails are the clearest test of in-image text: three or four huge words, high contrast, 16:9, and a face or object that reads at 200 pixels wide. Publishing daily means a fresh set every day, and A/B testing means several per video. Nano Banana 2 at $0.05 makes per-video thumbnail testing routine rather than a luxury.

Charts, infographics and text-heavy visuals

Infographics are where you should test Nano Banana 2 against your own standard rather than trust a claim. It handles a title, axis labels and a short legend well; it starts to slip when a chart carries twenty small labels that all have to be correct. Generate at 2K — same $0.05 as 1K — and the extra pixels buy you legibility for free, which is the single most affordable quality upgrade available on this model.

Price is the visible reason; access is usually the real one. Calling Nano Banana 2 here needs no Google Cloud project, no billing account, no quota request and no organisation verification — you sign up, top up, and the same key that runs Nano Banana 2 also runs Claude, GPT, Gemini text models, Seedance video and the rest of the catalogue. The endpoint is reachable from mainland China directly, which is the reason a large share of our Nano Banana 2 traffic is here at all.

Operationally: failed generations are never billed, spending is pay-as-you-go with no monthly floor, and every call is recorded with its prompt and result in the console so a cost question has an answer. Where a model has more than one upstream channel we fail over automatically rather than returning an error, which matters more for image endpoints than people expect.

If unit price is the only thing you are optimising, there is a more affordable sibling worth knowing: the same Gemini 3.1 Flash engine is also served on a budget channel at $0.025 per image, capped at 1K. Nano Banana 2 at $0.05 buys the 2K option and the primary channel's stability; for large batches where 1K is enough, the budget channel is the better deal and we would rather you knew that.

Coming from Google directly, the migration is a base URL, a key and a parameter rename: aspect ratio and resolution become aspect_ratio and resolution (1k / 2k / 4k), reference images go in image_urls, and the response carries a hosted URL. Prompts do not need rewriting — it is the same Gemini 3.1 Flash Image model underneath, so Nano Banana 2 output matches what you were already getting, at a lower line item.

Coming from another aggregator, the questions worth asking are the ones that decide whether a low headline price is real: does it bill failed generations, does it silently downgrade 2K requests to 1K, does it keep your result files longer than a few days, and does it publish which upstream serves the model. We bill only successful images, never downgrade a resolution you asked for, keep results for 30 days, and say plainly that Nano Banana 2 is Google's Gemini 3.1 Flash Image.

The lowest-risk way to move is to run both for a day: same prompts, same aspect ratios, both at 2K, and compare the two sets on your own subject matter rather than on a benchmark grid. Nano Banana 2 is affordable enough that this comparison costs a few dollars, and it settles the question better than any table on a landing page — including this one.