Title card reading Next Nano Banana on Gemini 4, with the subtitle a rumour from 18 November 2026, three chips reading Publisher: no model ID, Lineage: not a documented capability and Status: unreleased, and two outlined cards labelled Nano Banana 2.1 and Gemini 4 joined by a dashed arrow, with a footer noting that Google has not announced a successor
Guides & Insights

Next Nano Banana on Gemini 4: A Rumor That Collides With Google's Own Spec Sheet

Author

Gideon Frost

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

On the morning of 18 November, TestingCatalog posted that "the next Nano Banana, based on Gemini 4" was in preparation for release, listing it alongside three other rumoured Gemini names — Neon, Helix and Krypton. Take the claim at face value and you get a tidy story: Gemini Nano Banana 2.1 has led Google's image line for six weeks, and a successor built on the Gemini 4 generation would be the obvious next step. Read Google's own documentation and the story develops a crack. Gemini 4 Carbon, generally available since 10 November, lists image generation as not supported. So does Gemini 4 Hydrogen, out since 12 November, and so does Gemini 4 Argon, which has been on sale since 13 October. Every Gemini 4 model Google ships today is text-in, text-out. So is Gemini Nano Banana 2.1's predecessor Nano Banana 2, still callable as gemini-3.1-flash-image despite being deprecated in October. Nothing about that makes the rumour false. It does mean the interesting question is not "which model" but "which layer" — and that question has a checkable answer in the model reference. This is a what-we-know-so-far piece: no model ID, no price, no card, no date has been published for the supposed successor, and everything below is labelled accordingly.

What the post actually says, and what it does not

The post is four sentences and two images. The rumour inventory is Neon, Helix and Krypton — the last of which we have already covered as an unreleased codename with no model card behind it — plus "a potential Nano Banana upgrade." The only sentence with any specificity is the last: "The next Nano Banana, based on Gemini 4 is the one that was in a preparation for the release." The screenshots, whatever they show, carry no API identifier, no version number, no pricing row and no date. The quoted post beneath it is a single word: "Gemini."

That is thin even by leak standards. There is no internal string to decompose, as there was with the "gempix" naming report from October. There is no artefact that can be matched against a changelog. There is a claim about lineage and a claim about timing, and both are the kind that a rumour account can make from a staging console without being wrong or right for weeks. TestingCatalog has a reasonable record on spotting things early — that is the whole reason the post is worth reading — but "in preparation" is a statement about intent, not about a shipped artifact, and it should be handled as such.

Why "based on Gemini 4" is the load-bearing phrase

Strip the branding and "based on Gemini 4" is doing one of two very different jobs.

Read one: the successor inherits image output from a Gemini 4 model. That reading dies on contact with the spec sheet. Open the model pages for Gemini 4 Carbon, Gemini 4 Hydrogen and Gemini 4 Argon and each one lists Image generation as not supported, alongside a text-only output list. Carbon's page is explicit that it returns text and nothing else, at a 1,048,576-token context and up to 65,536 output tokens. There is no Gemini 4 model anywhere in Google's reference that emits an image, which means there is no Gemini 4 image model to inherit from — only a Gemini 4 generation to build on top of.

Read two: the image model uses a Gemini 4 weights family as its understanding layer and keeps a dedicated image head, the way Gemini Nano Banana 2.1 keeps a Flash-shaped decoder on top of its own backbone. This reading is entirely plausible, and it is the one that would actually matter to a user. The improvements a stronger backbone buys are the ones image models compete on now: prompt adherence, multi-turn consistency, legible text inside the render, and knowing what a "PlayStation controller" or a "windmill at dusk" is supposed to look like before a single token is sampled. Those are precisely the four things Google claimed to improve when it shipped Nano Banana 2.1 on 6 October. A Gemini 4 backbone is the obvious way to improve them again.

Note what this reading also implies: a successor would not be a new capability, it would be the same capability with a better brain. That is a much smaller story than "Gemini 4 can draw now," and it is the one the evidence supports.

Where the current ladder actually stands

To judge whether a successor is plausible at all, it helps to know exactly how much room is above Nano Banana 2.1. The Artificial Analysis image arenas put numbers on it, read on 18 November 2026. Text-to-image, higher is better:

• Gemini Nano Banana 2.1 — 1176.06 Elo, confidence interval 1164.06 to 1188.06

• Nano Banana 2 — 1143.40 Elo, the model Google deprecated on 6 October

• Nano Banana Pro — 1118.99 Elo, still the most expensive Google image model on the board

• Nano Banana 2 Lite — 1109.98 Elo

• GPT Image 2.6 Sunburst — 1260.66 Elo, the top row in the same arena

• MAI-Image-2.7 — 1214.39 Elo, second among non-OpenAI entries

Image editing, same source and same date:

• Nano Banana 2.1 — 1152.31 Elo

• Nano Banana 2 — 1125.57 Elo

• Nano Banana Pro — 1110.54 Elo

• Nano Banana 2 Lite — 1052.35 Elo

• MAI-Image-2.7 — 1215.17 Elo, first

• GPT Image 2.6 Sunburst — 1210.84 Elo, second

Two facts fall out of that. First, Nano Banana 2.1 is genuinely Google's strongest image model on both arenas — the tier ladder inside the Nano Banana family is intact, with Pro below the Flash-class model on both axes, which is an unusual inversion and a real reason to expect a Pro refresh rather than a Flash one. Second, Google is not close to the top. The gap is 84.60 Elo on text-to-image and 62.86 on editing. Those are not noise; the confidence intervals do not come near overlapping. Any successor has an obvious place to go, and a bigger backbone is the cheapest route to it.

Two-column scoreboard titled Gemini Nano Banana 2.1 vs GPT Image 2.6 Sunburst, with the Nano Banana column reading text-to-image Elo 1176.06, image-editing Elo 1152.31, price 33 dollars 60 per 1,000 images, Elo per dollar 35.0, API id gemini-nano-banana-2.1 and image editing yes, against a GPT Image column reading 1260.66, 1210.84, 210 dollars 72, 6.0, gpt-image-2.6-sunburst and yes, with a footer citing Artificial Analysis image arenas on 18 November 2026

The price bar is the harder constraint

A new image model is a pricing decision before it is a quality decision, and this is where the Gemini 4 rumour gets awkward in a way the quality argument does not.

• Nano Banana 2.1 — $30 per million image output tokens, which lands at $0.0336 per 1K image; Artificial Analysis lists it at $33.60 per 1,000 images

• Nano Banana 2 — $60 per million image tokens; Artificial Analysis lists it at $67 per 1,000 images

• Nano Banana Pro — $120 per million image tokens; Artificial Analysis lists it at $134 per 1,000 images

• GPT Image 2.6 Sunburst — $210.72 per 1,000 images on the same board

Nano Banana 2.1 is the cheapest row in Google's line and one of the cheapest near the top of either arena. Its Elo-per-dollar figure — roughly 35 points of text-to-image Elo per dollar of list price — is about six times GPT Image 2.6 Sunburst's. That ratio is the product. A successor that runs a Gemini 4-class backbone and charges image-token rates to match would break it: Gemini 4 Carbon is listed at an introductory $2.00 in and $10.00 out per million tokens through 30 November, and Hydrogen at $1.50 in and $7.50 out. A Gemini-4-backed image model that inherited those economics per image token would be several times the price of the model it replaces, for an unknown quality gain.

Which is the test to apply when the announcement lands. If the successor arrives at or near $0.0336 per 1K image, the "based on Gemini 4" claim is about architecture and the pricing story is intact. If it arrives at Pro-level prices, the claim is about positioning and the Flash-class users it is nominally replacing have been quietly moved up a tier.

What would make this real, and what would not

The falsifiable list is short, and none of it exists on 18 November.

• A new identifier in the Gemini API model reference — the current list runs gemini-nano-banana-2.1, gemini-3.1-flash-image, gemini-3.1-flash-lite-image, gemini-3-pro-image and gemini-2.5-flash-image, and nothing Gemini 4-shaped that emits an image

• A changelog line with a GA date and a price, the way Nano Banana 2.1 got one on 6 October

• A model card whose capability row for image generation reads supported rather than not supported

• A row in an independent arena, which is the only place a quality claim can be checked without Google's involvement

• A pricing row, since every image model Google has shipped arrived with one

• A shutdown date for something it replaces — worth noting that Nano Banana 2 was deprecated on 6 October with no shutdown date announced, so the older Flash image model is still callable and still billed at twice the new model's rate

What would not make it real: another post from the same account, a screenshot with a blurred identifier, or a codename appearing in a framework's supported-models list. Engine work and model releases look identical from the outside until a price appears, and this pipeline has been burned by that distinction before.

Screenshot of Google's Gemini API model reference page in English, headed All Gemini 3 models with Model and Endpoint columns, listing the Nano Banana rows gemini-nano-banana-2.1, gemini-3.1-flash-image, gemini-3.1-flash-lite-image and gemini-3-pro-image among the Gemini 3 text endpoints, with no Gemini 4 model listed at all

What you can call while this is speculative

The practical question a rumour raises is not what to switch to, it is what to keep working while you wait. On OrcaRouter, the Gemini image models currently routed are Nano Banana 2 as google/gemini-3.1-flash-image-preview, Nano Banana Pro as google/gemini-3-pro-image-preview, the original gemini-2.5-flash-image, the Imagen 4.0 family, and the Gemini 4 text models gemini-4-carbon and gemini-4-argon for the reasoning side of an image pipeline. Gemini Nano Banana 2.1 is not on that list — it is not routed yet, and nothing here should be read as saying otherwise.

Screenshot of the OrcaRouter model page for Nano Banana 2 (Gemini 3.1 Flash Image Preview) in English, showing the identifier google/gemini-3.1-flash-image-preview, the byline by Google dated 2026-02-26, a listing price of 0.15 dollars, input of image plus text and output of image plus text, and buttons to get the API, try it in the playground and use it via API

That split is usable. A pipeline that captions or critiques images with a Gemini 4 model and generates them with a routed image model keeps the two layers independent, so a successor shipping or not shipping changes one line of configuration rather than the architecture. Routing the generation call across more than one image model on the same key — with automatic failover between them — is the cheap version of the same insurance, and because the platform passes provider list prices through at 0% markup, the per-image arithmetic above is the per-image bill.

For anyone who specifically wants Nano Banana 2.1's price and quality today, the vendor's own API and several third-party platforms carry it. What the next six weeks decide is whether that stays the answer.

The open question

Three codenames in one post is a lot of signal for a Tuesday morning, and the least specific of the three is the one with commercial consequences. Neon and Helix have no artefacts at all; Krypton has demos and a token count but no identifier. The Nano Banana claim has neither, and it is the only one that implies a price change for a model people are already paying for.

The tell will not be a screenshot. It will be a row in the model reference with a date next to it and a number at the end. Until then, the honest summary is the one the spec sheet gives: Gemini 4 is a text generation. A Nano Banana built on it would be a better-reader of prompts, not a new kind of renderer — and the interesting part is whether Google can do that without doubling the price of the cheapest good image model it has.