Nano Banana 2.1: The Gemini Image Model, Its Family, and What It Costs

Nano Banana 2.1 is the image generation and conversational editing model the vendor serves under the API identifier gemini-nano-banana-2.1. The vendor's own documentation gives it the product name Gemini Nano Banana 2.1 and calls it an update to Nano Banana 2, whose identifier is gemini-3.1-flash-image. Two other models bracket it in the same family: Nano Banana 2 Lite (gemini-3.1-flash-lite-image) below it, and Nano Banana Pro (gemini-3-pro-image) above it. That is the shape of the thing — a three-tier image family with a fourth, older member still documented — and most of the confusion around the name comes from the fact that the product names and the endpoint strings were never made to match. This page is the reference for the model itself: what it is, who publishes it, what you can put in and get back, what the vendor claims, and how you reach it today. Every figure below is quoted from the vendor's own model card, image generation guide, pricing page and changelog as they stand on October 13, 2026, and each one is labelled vendor-reported; the single independent source used is named where it appears.

What it is, and why the identifier is the part that matters

The vendor's model directory lists gemini-nano-banana-2.1 with the status "New Stable" and the description "High-efficiency image generation and editing model with improved visual quality, typography, and multi-turn consistency." The image generation guide positions it more plainly still: it is "an update to Nano Banana 2, serving as the primary high-efficiency workhorse model for image generation and conversational editing."

The naming is the first practical problem a reader hits, so it is worth stating flatly. In this family the product name is not derivable from the identifier:

• Nano Banana 2.1 — product name "Gemini Nano Banana 2.1", API id gemini-nano-banana-2.1, the current high-efficiency workhorse.

• Nano Banana 2 — API id gemini-3.1-flash-image. No version number in the identifier at all, which is why "Nano Banana 2" and "gemini-3.1-flash-image" are the same model.

• Nano Banana 2 Lite — API id gemini-3.1-flash-lite-image. Google's guide calls it the efficiency specialist of the family, built for ultra-low latency and high-volume interactive use.

• Nano Banana Pro — API id gemini-3-pro-image, described as a professional design engine with a reasoning core, aimed at complex layouts and precise text rendering.

If you are wiring something up, the identifier is the reliable handle and the marketing name is not. This is not a stylistic point: Google's changelog moved gemini-3.1-flash-image to deprecated status on the same day it released gemini-nano-banana-2.1, and the deprecation entry does its work in identifiers rather than in marketing names.

Who publishes it

Google does, and it publishes it as a hosted API model rather than a downloadable one — there is no checkpoint to fetch and no licence to read, because the weights are not released. The authoritative material is Google's own Gemini API documentation set: the model directory and per-model card on ai.google.dev, the image generation guide, the pricing page, and the Gemini API changelog. Where this page gives a number, that is the source, and it is a vendor source.

The family relationship matters more than the publisher line, though, because it decides what you can actually choose between. Against Nano Banana 2, its immediate predecessor, 2.1 is the generation that Google is steering new projects toward; against Nano Banana 2 Lite, 2.1 buys a higher ceiling at a matching entry price, since both are billed at $0.0336 for a 1K image (vendor pricing, standard tier). Lite gives that up in three documented ways: it supports only 1K output, it has no Search grounding, and Google's guide says it is "not optimized for multiple reference inputs or multi-turn sequential editing." If your workload is a single image in and a single image out at volume, Lite is the cheaper engineering choice; if you need 2K or 4K, several references, or a conversation with the image, Lite is not a candidate at all.

What it accepts and what it returns

This is where the model is wider than its name suggests, and it is worth reading off the vendor card rather than inferring from the word "image". Google's model card for gemini-nano-banana-2.1 lists the accepted input data types as Text, Image, Video and PDF, and the output types as Image and Text. So the input side takes a prompt, one or more images, video, or a PDF document — not text alone — and the output side can return text alongside the image, which is what makes conversational editing possible: the model can answer about the picture it just made or changed.

The rest of the envelope, all vendor-reported from the model card, the guide and the changelog:

• Output resolutions — 1K, 2K and 4K. The 512px (0.5K) tier that Nano Banana 2 offers is explicitly "not supported on Gemini Nano Banana 2.1", so on this model the floor is 1K. Parameters must use an uppercase K; Google's guide states that lowercase values such as 1k are rejected.

• Aspect ratios — the guide now publishes a per-model table and 2.1's column carries fourteen: 1:1, 1:4, 1:8, 2:3, 3:2, 3:4, 4:1, 4:3, 4:5, 5:4, 8:1, 9:16, 16:9 and 21:9. That includes the wide and panoramic ratios at 1:4, 4:1, 1:8 and 8:1, which the guide also credits to Gemini 3.1 Flash Image. Lite lists the ten conventional ratios only.

• Reference images — "Mix up to 14 reference images to produce the final image," per the guide, and the guide's table now assigns the budget per model, which it did not do legibly when this page was first drafted. For the column headed "Gemini Nano Banana 2.1 & Gemini 3.1 Flash Image" the split is up to 10 object images with high fidelity and up to 4 character-consistency images; the guide's prose gives the same numbers, saying 2.1 supports "character resemblance of up to 4 characters and the fidelity of up to 10 objects in a single workflow." Style references are listed for Gemini 3 Pro Image only, so do not plan on them here.

• Token limits — 131,072 input tokens and 32,768 output tokens, per the model card. The card states no knowledge cutoff for this model.

• Thinking — supported, and the guide is explicit that on the Gemini 3 image models it is "enabled by default and cannot be disabled in the API." The model generates up to two interim images to test composition and logic, and those thought images are visible in the backend but not charged. A thinking_level can be set in the generation config; the guide's own example passes high.

• Grounding — Search grounding is listed as supported on the model card, and the changelog credits 2.1 with Google Search grounding. Nano Banana 2 Lite does not have it.

• Watermarking — "All generated images include a SynthID watermark." That applies across the family and is not optional.

• What the card marks unsupported — audio generation, caching, code execution, file search, function calling, Google Maps grounding, the Live API, structured outputs and URL context. Batch API is supported; Flex and Priority inference are not.

Text rendering and editing, as documented

Two documented behaviours separate a model like this from a plain text-to-image endpoint, and Google describes both in the image generation guide rather than in a benchmark.

The first is typography. The guide lists the capability as "Advanced text rendering: Capable of generating legible, stylized text for infographics, menus, diagrams, and marketing assets," and the changelog credits 2.1 specifically with improved text rendering. The model card echoes it in the word "typography". Worth being precise about what that is: a description of a capability, not a measured accuracy rate. Google publishes no character-error figure, no legibility score and no comparison against its own earlier model on a text-rendering task, so anyone quoting a number for how well 2.1 renders letters is quoting something other than Google.

The second is editing, which the guide treats as the model's conversational mode. On an edit, "the model will match the original image's style, lighting, and perspective" — the edit is applied inside the existing frame rather than producing a new picture in the same genre. For inpainting, the instruction is narrower still: change only the element you name and keep "the original style, lighting, and composition." The changelog's headline claim for 2.1 on this axis is improved multi-turn character consistency, which is the property that decides whether an object or a person survives five sequential edits looking like themselves. That claim is vendor-reported and, as of today, unreproduced outside Google.

What Google claims versus what has actually been measured

Google's changelog entry for October 6, 2026 says 2.1 "maintains Flash-level speed and cost efficiency" while adding "significant improvements in visual quality, prompt adherence, multi-turn character consistency, text rendering," and wide and panoramic aspect ratios. The model directory adds "improved visual quality, typography, and multi-turn consistency." Those are the vendor's words, they are directional rather than quantified, and no vendor-run benchmark accompanies them.

Three things follow, and they are the honest state of the evidence rather than a hedge.

First, Google publishes no benchmark figures of its own for Nano Banana 2.1 — no Elo, no win rate, no score of any kind on its documentation pages. If a number about this model's quality is presented as Google's, it is a number Google has not published.

Second, there is no vendor head-to-head between Nano Banana 2.1 and Nano Banana 2 at matched settings. Google claims improvements from one generation to the next; it has not published a paired evaluation at matched prompts, matched resolutions and matched effort that would let a reader weigh them. Anyone implying a measured gap between the two is implying something nobody has run.

Third, independent evaluation does exist, and the harness is worth naming so the reader can judge it: Artificial Analysis runs Nano Banana 2.1 in its Image Arena, which covers both text-to-image and image editing and publishes the votes behind its scores. That is the third-party source to use for placement. The specific rank and Elo this model holds there is a different question from the one this page answers, and it is computed on a schedule this page has no control over — check it at the source rather than against a figure copied from anywhere, including from us.

How to get access today, and what it costs

Access is through Google's own API. You call the Gemini API with the model identifier gemini-nano-banana-2.1; that identifier is the stable one, and it is what the changelog's deprecation entry for gemini-3.1-flash-image names as the migration target. There is no self-hosted route, no weight download and no community checkpoint, because none of the image family is released that way.

Pricing, all of it from Google's pricing page in the standard tier and all of it vendor-reported:

• Tokens — input $1.50 per 1M tokens for text, image and video; output $7.50 per 1M tokens for text and thinking, and $30.00 per 1M tokens for images.

• Per image — $0.0336 at 1K, $0.0504 at 2K and $0.113 at 4K. Google also publishes a worked equivalence for the 1K case: a 1024×1024 output consumes 1,120 image tokens.

• Batch — half the standard rate: $0.75 per 1M input and $15.00 per 1M image tokens, which works out to $0.0168 / $0.0252 / $0.0567 per 1K / 2K / 4K image, in exchange for a turnaround of up to 24 hours.

• Grounding — no separate line item on the image itself. Google bundles Web and Image Search grounding with the Gemini 3 image models at 5,000 free search requests per month, shared across all Gemini 3 and newer models, then $14 per 1,000 requests; retrieved context is not billed as input tokens.

• Free tier — the pricing page's Free Tier column reads "Not available" for every figure above. There is no free allowance for image output on this model. If one appears, it will appear there and nowhere else.

• Tier comparison within the family — Nano Banana 2 Lite is $0.0336 per 1K image, identical to 2.1's 1K rate, but it publishes no 2K or 4K price because it does not do those sizes. Nano Banana 2 (Gemini 3.1 Flash Image) is $0.067 per 1K and $0.101 per 2K, so 2.1 is cheaper at every size they share. Nano Banana Pro (Gemini 3 Pro Image) is $0.134 per 1K or 2K image and $0.24 at 4K, roughly twice 2.1's rate through 2K and just over twice at 4K.

• The legacy member, for completeness — the original Nano Banana (gemini-2.5-flash-image) is deprecated and the pricing page carries a shutdown notice dated October 2, 2026. That model is not a tier below 2.1 in any live sense; the family's entry point now is Nano Banana 2 Lite.

One figure inside that set does not reconcile, and it is worth saying so rather than picking the tidier number. Google's pricing page states that a 4K image on this model consumes 3,780 tokens and quotes $0.113 per 4K image; those two agree at $30 per 1M image tokens. The image generation guide's per-model table gives the same 1K and 2K token counts as the pricing page (1,120 and 1,680) but lists 2,520 for 4K, which at the same rate would be about $0.0756. Both are Google pages and they disagree. The price is the figure to hold onto, because it is the one on the page that bills you, and the discrepancy is exactly the kind of thing to re-check at the source before you build a cost model on it.

One more line a reader may want, stated as what it is: Nano Banana 2.1 is not on OrcaRouter's live model list today, so this page makes no claim about calling it here. What is on that list is its direct predecessor under the identifier google/gemini-3.1-flash-image-preview, which is the model whose strings migrate to 2.1. OrcaRouter passes provider list price through with 0% markup on the models it routes, so a vendor price change on a routed model lands here the same day rather than on a delay — which is the property to check for on this family as Google keeps re-tiering it.

The short version

Nano Banana 2.1 is Google's mid-tier Gemini image model: text, image, video and PDF in; image and text out; 1K, 2K and 4K; up to 14 references split ten objects and four characters; Web and Image Search grounding; SynthID on every output; thinking on by default; $0.0336 per 1K image. It sits above Nano Banana 2 Lite, which matches its 1K price but stops at 1K and drops Search grounding, and below Nano Banana Pro, which is roughly twice the rate and the expensive route to complex layouts and precise text. Its identifier, gemini-nano-banana-2.1, is the only name that will not mislead you — and the quality claims attached to it are Google's words, unreproduced. The model page is the place to learn what it is; if you are choosing it on a quality difference from Nano Banana 2, the only place to see that difference weighed is the independent arena, not the vendor's release note.