Hero title card for the article 'Nano Banana 3 Pro — LEAK REPORT', carrying an 'UNVERIFIED — NO GOOGLE ANNOUNCEMENT' badge, the kicker 'IMAGE-MODEL SIGNAL REPORT', the title 'Nano Banana 3 Pro — LEAK REPORT', the subtitle 'Reported internal name: Gempix 3 Pro. Reported foundation: Gemini 4 Argon. Reported window: November-December 2026.', three chips reading 'Source: TestingCatalog', 'Reported Oct 12, 2026' and 'Status: in development, unreleased', and two stat cards reading 'Gempix lineage — the third use of Google's internal image-model string' and 'Gemini 4 Argon — announced Sept 30, 2026, still not callable from the Gemini API'. The OrcaRouter logo is composited in the bottom-right corner.
Guides & Insights

Nano Banana 3 Pro Leak: the Gempix 3 Pro String Sitting in Google's Image Pipeline

Author

Magnus Corvin

Date Published

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

Nano Banana 3 Pro has not been announced. What exists is a string — Gempix 3 Pro, the internal designation TestingCatalog reported on 2026-10-12 after finding it referenced in the vendor's own image-recitation plumbing — and a lineage that makes the reading obvious: Gempix was the pre-launch name for the original Nano Banana, GEMPIX 2 turned up in the consumer UI a week before Nano Banana 2 shipped, and a third generation carrying "Pro" would be the successor to Nano Banana Pro, the gemini-3-pro-image-that-was that the vendor has since replaced underneath its consumer lineup with Gemini Nano Banana 2.1. Every claim below is a report, not a release. There is no model card, no identifier, no price, no benchmark, and no date.

That last part is the whole point of writing this now. The value of a leak piece is not the leak; it is being able to tell, in November, whether the thing you are looking at is the thing that was predicted. So this page is built backwards from that: what the report actually says, what the Gempix naming convention has meant twice before, what "based on Gemini 4" can and cannot mean given where Gemini 4 currently sits, what the image-recitation detail is really evidence of — and what you can call today if the answer to "should I wait" turns out to be no.

What the report actually says

The claim is short enough to state in one breath. The company appears to be preparing another image-generation model; references to a string called Gempix 3 Pro have been spotted; it previously used the Gempix 2 designation for Nano Banana 2 Pro; Gempix 3 Pro is therefore read as Nano Banana 3 Pro, an upgrade to the Pro image tier; and the new model is said to be based on Gemini 4 Argon, the Gemini 4 model announced on 2026-09-30. The report also notes that Gempix 3 Pro has been added to its existing image-citation functionality, and that as of publication there are no confirmed plans for a public release, with November-December described only as a plausible window based on its prior release cadence.

That is the entire public record. It is one outlet, one unnamed observation, and an inference from a naming pattern — not a screenshot of a model picker, not a discovery-card leak like the one that telegraphed Nano Banana 2, and not a vendor page with a date on it. TestingCatalog has a very good track record on this specific family, which is why the report is worth a page at all, but a track record is not a confirmation and this article does not treat it as one.

Extracted to its verifiable minimum:

• Name claimed — "Gempix 3 Pro", an internal designation, not a product name and not an API identifier.

• Mapped product name — Nano Banana 3 Pro, by analogy with Gempix 2 → Nano Banana 2 Pro.

• Reported foundation — Gemini 4 Argon, announced 2026-09-30.

• Publication — TestingCatalog, 2026-10-12.

• Corroboration — none found. No second outlet, no vendor acknowledgement, no model card, no pricing entry, no identifier in any vendor API surface.

• Confirmed release plan — none. The November-December window is the outlet's inference from the vendor's own release history, and it is labelled that way at the source.

A what-we-know-so-far infographic titled 'Nano Banana 3 Pro — the leak, so far', subtitled 'Everything currently on the public record, separated from the guesses', listing six rows: Internal string 'Gempix 3 Pro', Reported product name 'Nano Banana 3 Pro', Source 'TestingCatalog, Oct 12, 2026', Google confirmation 'none', Model card, ID or price 'none', and Suggested release window 'November-December 2026 — inference, not a plan'. Footer: 'Single-source report — unverified. Nothing has shipped. Reported foundation (Gemini 4 Argon) is itself announced but not yet callable.' The OrcaRouter logo is composited in the bottom-right corner.

Why "Gempix" is the part worth believing — and the part that has failed before

Gempix is not a new word, and the way it has been used twice already is the strongest signal in this report. It is also the weakest, because one of those two uses turned out to be a codename attached to a later stage than the reporting implied. Both halves matter.

The first appearance is well documented and dated. In August 2025, Gempix showed up inside Whisk, Google's experimental image tool, hidden in a "Precise Reference" toggle offering three modes — Default, GEM_PIX and R2I. At the time it was read as a new image-editing model built on top of Imagen 4, possibly tied to the Pixel event on August 20, and publicly available Whisk still ran on Imagen 3 while trusted testers had Imagen 4. That reading was wrong about what the string was and right about what it meant: Gempix turned out to be the internal name for what Google shipped publicly as Nano Banana, the model that arrived as Gemini 2.5 Flash Image.

The second appearance is the one this leak rests on. In early November 2025, pre-release announcement cards appeared in the consumer web interface labelled "GEMPIX2" — not a model picker, not an API string, but a card in the consumer product, which is what Google puts in front of users shortly before a rollout. Nano Banana 2 followed. That is the precedent that converts "we found a string" into "something is close", and it is why a Gempix 3 Pro sighting is more than trivia.

Except for the part that the current reporting gets slightly loose. The 2025 article mapped GEMPIX 2 to "Nano Banana 2 Pro", and the string that shipped publicly, gemini-3-pro-image-preview, was indeed the Pro build of Nano Banana 2. But the model the public met first, and the one Google still calls Nano Banana 2, is the Flash-tier gemini-3.1-flash-image. So a Gempix 3 Pro sighting tells you a Pro-tier internal build exists. It does not tell you it will be the first Nano Banana 3 you see, and it does not tell you the public name will be "Nano Banana 3 Pro" rather than something the marketing team chooses later. Internal strings and product names in this family have never matched, which is the single most reliable thing about it.

One more detail in the report deserves separating from the headline. TestingCatalog notes that Gempix 3 Pro has been added to Google's image-citation system — the mechanism that detects when a generated image closely matches a reference image from somewhere else and blocks the generation; Google's API documents the stop reason as IMAGE_RECITATION. That is not evidence of capability, and it is not a benchmark. It is evidence of a different and quite useful kind: eligibility lists like a recitation filter are driven from a registry of models that are meant to serve traffic, and a codename does not get added to one for fun. It is the sort of plumbing that fills in ahead of a launch, which puts the observation in the same category as the announcement cards of November 2025 rather than in the category of an idle string in a JS bundle.

A three-row lineage card titled 'The Gempix string, three sightings', subtitled 'One internal codename, two shipped models, one open question', listing: August 2025 — 'GEM_PIX' appears in a Precise Reference toggle inside Whisk; read at the time as an Imagen 4 editing model; what ships publicly is Nano Banana, the model already replaced in Google's consumer lineup by Gemini Nano Banana 2.1. November 2025 — 'GEMPIX2' announcement cards appear in the Gemini web interface about a week before rollout; Nano Banana 2 follows, with the Pro build of the same generation shipping publicly as Nano Banana Pro (gemini-3-pro-image). October 2026 — 'Gempix 3 Pro' is reported inside Google's image-recitation path and in the picker of a Google Labs experiment, reported Oct 12, 2026; no announcement, no model card, no identifier, no price, no date. Footer: 'First two sightings per TestingCatalog and Google's public rollout; the third is a single report — unverified. Internal strings and public product names in this family have never matched.' The OrcaRouter logo is composited in the bottom-right corner.

"Based on Gemini 4" is doing more work than it can carry

The claim that Nano Banana 3 Pro is based on the Gemini 4 model is the most seductive sentence in the report and the one that needs the most care, because "Gemini 4" currently names less than it sounds like.

What exists today is Gemini 4 Argon: announced by Google on 2026-09-30, real enough to have a first-party family page and a nineteen-row vendor benchmark table, and measured independently by Artificial Analysis at an Intelligence Index of 52.56 — but with no model identifier in the Gemini API. The changelog's newest entries are October 8, 2026, and they add deep-research previews; the Gemini Nano Banana 2.1 entry sits at October 6. Nothing in the gemini-4 namespace appears anywhere in Google's API surfaces. As the blog's pillar page on the name puts it, Argon is announced and benchmarked but not callable, and "Gemini 4" on its own is attached to nothing.

So the composition claim has to be read narrowly. It cannot mean that Gempix 3 Pro is a request-time wrapper over an Argon endpoint — there is no Argon endpoint. It can only mean that the image model descends from the Gemini 4 generation of Google's model work rather than from the Gemini 3.1 generation that the current image family is built on. That distinction is real and it is the most interesting thing in the leak, because it is the difference between a point release and a generational replacement.

Read it out against where the lineup actually stands. Nano Banana Pro is the Pro-tier image model built on Gemini 3 Pro. Nano Banana 2 is the Flash-tier image model built on Gemini 3.1 Flash. Nano Banana 2.1 is the high-efficiency model Google shipped into general availability on 2026-10-06 — and on the same day, it deprecated gemini-3.1-flash-image, the Nano Banana 2 the pricing page now calls the previous-generation high-efficiency workhorse. That is a family in motion: three generations of Gemini underneath it within twelve months, a Pro tier that is now the *older* of the two Pro-family names still in circulation, and a documented instruction telling new projects to use 2.1 rather than 2.

Against that backdrop a Gempix 3 Pro on a Gemini 4 foundation is not an incremental drop. It would be the first image model in the family built on a new base model, and — if Google follows what it did with Nano Banana 2 — it would be the leading edge of a whole Nano Banana 3 generation that eventually produces a Flash-tier sibling on the same base. That is the reading the report supports, and it is a forecast, not a finding.

What a leak like this can and cannot tell you

It is worth being mechanical about this, because "Google is preparing X" reports get consumed as if they were spec sheets.

What this report can tell you, with reasonable confidence: that Google's image pipeline contains a Pro-tier codename one generation past the one that shipped, that the codename is plugged into a production-side content-protection path, and — if the Argon claim holds — that the company intends the next image generation to ride a Gemini 4 base rather than extending the 3.1 line. Those are structural facts about direction, and direction is genuinely useful for anyone deciding whether to build on this family now or wait.

What it cannot tell you, and what no leak of this shape ever can: the resolution ladder, whether the wide and panoramic ratios that Nano Banana 2.1 added survive into a Pro build, the reference-image budget, whether the model reasons before it renders and at what cost in latency, what Google search-grounding will look like one generation on, and the price. Every one of those is a number someone will attach to this name in the next eight weeks, and none of them is in the record today. The one thing the leak does imply about capability is directional rather than numeric: a model built on a new base is the only mechanism by which this family gets a step change rather than another tuning pass, and "another tuning pass" is exactly what the October 6 release already was.

There is a detail worth noting about how you will know the leak is real. Nano Banana 2 reached consumers on 2026-02-26, and Nano Banana 2.1 surfaced first as a selectable label in the Google Flow model picker on 2026-10-05, one day before the GA changelog entry. Consumer surfaces, not API docs, are where this family announces itself. If a third generation is genuinely coming in November or December, the first reliable sign will be a string in a picker or a card in the Gemini interface — not a blog post. And when that string appears, the question to ask is which of the release's two names it is.

What you can actually call while this is speculative

None of the models in this story are callable from an API today under the names the leak uses. Gemini 4 Argon has no identifier; Gempix 3 Pro is an internal string; Nano Banana 3 Pro is a product name nobody at Google has used in public. What is available is the current generation of the same family, and reading its price ladder is the fastest way to understand what a new Pro tier would have to justify.

• Nano Banana (Gemini 2.5 Flash Image) — the original, $0.30 per million input tokens and $30.00 per million output tokens on Google's list, roughly $0.039 per 1K image. Google's pricing page carries a shutdown notice for it dated October 2, 2026, and the changelog has since limited 2.5-series access to users who were already active on those models. It is a migration target in progress, not a foundation.

• Nano Banana Pro (Gemini 3 Pro Image) — $2.00 per million input tokens and $120.00 per million image output tokens, which Google equates to $0.134 per image at 1K or 2K and $0.24 at 4K. This is the tier the leak is about, and the reason the tier matters: through 2K it is roughly two and a half times the price of the current Flash-tier workhorse for the same nominal resolution.

• Nano Banana 2 (Gemini 3.1 Flash Image) — $1.00 per million input tokens, $0.067 per image at 1K or 2K and $0.12 at 4K. Deprecated as of 2026-10-06 with no shutdown date announced, which is a strange state for a model to be in: still documented, still priced, formally on the way out.

• Nano Banana 2.1 (Gemini Nano Banana 2.1) — $1.50 per million input tokens and $0.0336 per 1K image, $0.0504 at 2K and $0.113 at 4K, with a batch rate at half. It is the model Google is steering new projects toward, it is cheaper per image than everything above it except the Lite tier, and it is the practical baseline any Nano Banana 3 Pro has to beat rather than the Pro tier it nominally replaces.

That comparison is the uncomfortable one the leak does not make. A Nano Banana 3 Pro priced anywhere near Nano Banana Pro's $0.134 per 1K image would be four times the per-image cost of the model Google shipped eight days ago, on a resolution ladder where 2.1 already reaches 4K. Whatever a generation jump buys, the Pro tier's price is now set against a moving floor, and 2.1 moved it.

Much of that family is reachable today through a single OrcaRouter key, each model at the provider's own list price passed through with no markup, so the October 6 price change on Nano Banana 2.1 landed here the same day rather than after a billing cycle. Automatic failover matters more than usual in this particular family: with gemini-3.1-flash-image formally deprecated and no shutdown date published, a stack that pins a snapshot string has no scheduled break to prepare for, only an announcement that may not give much warning. Routing the call through one endpoint makes a model-string swap a config change, and when a Nano Banana 3 Pro eventually appears — or when it does not, and 2.1 turns out to be the answer for another six months — that is the difference between one line and one migration.

A screenshot of the English-language OrcaRouter model page for Gemini 3 Pro Image Preview, the Pro-tier image model the leak's successor would replace, showing the breadcrumb Home / Models, the heading 'Gemini 3 Pro Image Preview' described as 'a next-generation image generation and editing model', the vendor Google, a price note reading 'Priced per generated image (e.g., $0.24 for a 2K image)', input type Image + text, output type Image + text, a 65K-token context window and a Get the Gemini 3 Pro Image Preview API button.

How to tell, in six weeks, whether this was the real thing

The report's predecessor two and a half weeks earlier read a Flow picker label — Nano Banana 2.1 — as the next image model and inferred it was a Flash-tier half-step up from Nano Banana 2. The label was real and the launch followed within days, and the substantive change was something the label could not have carried: a per-image price roughly halved against the tier directly beneath it, plus the deprecation of the model it was cheaper than. That is the pattern to expect here, and the reason it is worth stating precisely is that it tells you what to watch for rather than what to hope for.

Three things would settle this leak, in descending order of how much they would settle.

First, a consumer-surface string. GEMPIX 2's public debut was a card in the Gemini web interface about a week before rollout; 2.1's was a selectable model label in Flow about a day before the changelog entry. If "Gempix 3 Pro" or a Nano Banana 3 name appears in a picker, a card or an announcement screen, the leak is confirmed and the timeline starts the same week — in which case expect whatever generation ships to have been measured by Artificial Analysis' image arenas within days, as happened to 2.1 on both boards.

Second, an identifier. Google's changelog dates its entries and has published one for every model in this family: the model directory and pricing page will carry a new entry, with a per-image price for each resolution tier, on the day it becomes generally available. A Nano Banana 3 Pro price line is the single highest-value piece of information anyone will get, because — as the 2.1 release showed — this family competes on price per image at least as much as on quality.

Third, and only if Argon resolves first: an Argon identifier in the Gemini API. The claim that the next image model rides on Gemini 4 is downstream of Gemini 4 becoming callable at all. If November and December pass with Argon still listed as announced-but-unavailable, the foundation claim is the part of this leak to disbelieve, not the codename.

What to do in the meantime depends on which side of the Pro tier you were going to buy. If your workload needs the Pro tier's layout and typography, nothing here changes your next six weeks: gemini-3-pro-image is what exists and what you can call, and a leak is not a migration plan. If your workload has been waiting for a cheaper Pro, the honest position is that Nano Banana 2.1 already did most of what you were waiting for — $0.0336 per 1K image, 4K support, search grounding, and the vendor's own deprecation of the model above it in the tier order. Waiting for a name that has been reported exactly once, with no date and no price, is a bet on a rumour when the alternative is a model that shipped on October 6.

Nano Banana 3 Pro is, for now, a string in someone else's pipeline. It is a good string — the Gempix convention has preceded two real launches — but it is not a product, it has no date attached by anyone who could actually set one, and the tier it would occupy has just been repriced downward by the release beneath it. Track it. Do not schedule around it.