Nano Banana 2.5 Flash on Google Flow: Essential Upgrade Risk

Nano Banana 2.5 has shown up where Google’s image models tend to appear first. On 24 September 2026, TestingCatalog, a publication that tracks unreleased AI features, posted that “Nano Banana 2.5 Flash traces have been spotted on Google Flow”, followed by a single word: “Soon?” Google has not announced the model, and the post included one screenshot but no specifications, price or date.

The sighting fits a pattern. For two weeks, testers on Image Arena have been comparing an anonymous image model codenamed spicy-mayo, which several of them believe is the next Nano Banana. Flow, Google’s AI creative studio, already uses Nano Banana 2 as its free default image model. It is the obvious place for a new Flash-tier image model to appear.

This article sets out what was spotted and what it does and does not prove. It covers how a Nano Banana 2.5 would fit Google’s current image models, what the spicy-mayo tests suggest, today’s prices and the 2 October shutdown of the original Nano Banana. It ends with what developers, creative teams and businesses should do while they wait.

Table of contents

What Was Spotted: Nano Banana 2.5 Flash Traces in Google Flow
Where Nano Banana 2.5 Fits in Google’s Image Line-Up
The Spicy-Mayo Tests Behind the Nano Banana 2.5 Rumours
Why Google Flow Is the Natural Home for Nano Banana 2.5
What Nano Banana 2.5 Has to Beat: Nano Banana 2 Today
Nano Banana 2.5 Pricing: What Current Rates Suggest
The 2 October Deadline That Will Not Wait for Nano Banana 2.5
What Nano Banana 2.5 Could Change for Creative Teams
What Nano Banana 2.5 Means for Businesses Building on Gemini
When Could Nano Banana 2.5 Launch?
Nano Banana 2.5 FAQ
References

What Was Spotted: Nano Banana 2.5 Flash Traces in Google Flow

The evidence so far is thin but specific. TestingCatalog is the source of the Flow sighting, and it has a long record of finding Google features in code before launch.

The TestingCatalog post

The post appeared on TestingCatalog’s Threads account on 24 September. Its full text reads: “GOOGLE 🔥: Nano Banana 2.5 Flash traces have been spotted on Google Flow. Soon? 👀” It carried one image. The same day, the account also reported that Google is testing a Task mode for the Gemini desktop app with Gemini Live support, which shows how much Google has in testing at once.

Nokiapoweruser wrote up the sighting the same day. It said that backend inspections of Flow showed references to “nano-banana-2.5-flash endpoints”. It did not publish those references, so treat that detail as reported, not confirmed. Nothing in the Gemini API documentation, updated on 24 September, mentions a Nano Banana 2.5 model.

What a trace usually means

A trace is a reference to an unreleased feature in a product’s code, such as a model name, a menu label or an unused setting. Leak trackers read Google’s web apps for these strings. A trace shows that engineers have wired something up. It does not show that the feature works, that it will ship or what it will be called.

TestingCatalog’s recent record on Google suggests that traces are worth taking seriously. On 7 May 2026 it reported code for an Agent Mode in Flow, and Google launched Flow Agent at I/O on 19 May, twelve days later. We have also covered earlier code traces pointing to Gemini 4 in the desktop app.

What the traces do not tell us

The more useful lesson comes from the last Nano Banana leak. In December 2025, TestingCatalog found code references to a model it called Nano Banana 2 Flash, codenamed Mayo. On 23 February 2026 it reported that Google had halted that model in favour of a newer one, Gemini 3.1 Flash Image. That model shipped three days later as Nano Banana 2. The traces pointed to a real product line, but both the leaked name and the model changed before launch. The same could happen to Nano Banana 2.5 Flash.

Claim about Nano Banana 2.5	Source	Date	Status
“Nano Banana 2.5 Flash” traces exist in Google Flow	TestingCatalog	24 Sep 2026	Reported, with a screenshot
Flow references “nano-banana-2.5-flash” endpoints	Nokiapoweruser	24 Sep 2026	Reported, no evidence shown
An anonymous Google image model, spicy-mayo, is on Image Arena	Several testers on X	9 Sep 2026	Observed by many testers
spicy-mayo is Nano Banana 2.5	Testers’ inference	9 Sep 2026 onwards	Unconfirmed
Partners can test it through Vertex AI, with three thinking levels	Posts relayed by aggregators	From about 18 Sep 2026	Unconfirmed
Google has announced Nano Banana 2.5	Google	None	Not announced

Where Nano Banana 2.5 Fits in Google's Image Line-Up

Nano Banana began as a nickname and became a brand. Knowing the family makes it clearer what a 2.5 release would add.

From Nano Banana to Nano Banana 2 Lite

The original Nano Banana is Gemini 2.5 Flash Image, which Google announced on 26 August 2025. It became popular for editing photos with plain instructions while keeping faces and objects consistent. Nano Banana Pro, built on Gemini 3 Pro Image, followed on 20 November 2025 and added studio-quality control and 4K output.

Nano Banana 2 arrived on 26 February 2026. It is Gemini 3.1 Flash Image, and Google described it as Pro-level quality at Flash speed. On 30 June 2026 Google added Nano Banana 2 Lite, or Gemini 3.1 Flash Lite Image, as its fastest and cheapest image model. So the current family has three live tiers and one legacy model.

The naming puzzle

The name “Nano Banana 2.5 Flash” contains two clues. The first is the Flash suffix. Google uses Flash for its fast, mid-priced tier, below Pro and above Lite, so a Flash label suggests an update to Nano Banana 2 rather than a new flagship. It also leaves room for a matching Pro or Lite version later.

The second clue is easy to misread. The original Nano Banana is Gemini 2.5 Flash Image, so a quick reading of “Nano Banana 2.5 Flash” could point back to the 2025 model. That reading does not hold. Google never marketed the original as Nano Banana 2.5, and the model is being retired. The version number follows the consumer name, as Nano Banana 2 did, not the Gemini engine behind it.

The condiment codenames

Google’s image models have carried food codenames in testing. TestingCatalog reported in December 2025 that Nano Banana Pro was known internally as Ketchup and the planned Flash model, which led to Nano Banana 2, as Mayo. The Arena model now being tested is called spicy-mayo. That fits a Flash-tier successor to Mayo, which is one reason testers link it to Nano Banana 2.5. It is still a guess, not a confirmation.

Model	API model ID	Announced	Tier	Status (24 Sep 2026)
Nano Banana	gemini-2.5-flash-image	26 Aug 2025	Flash (legacy)	Deprecated, shuts down 2 Oct 2026
Nano Banana Pro	gemini-3-pro-image	20 Nov 2025	Pro	Stable
Nano Banana 2	gemini-3.1-flash-image	26 Feb 2026	Flash	Stable, Google’s recommended default
Nano Banana 2 Lite	gemini-3.1-flash-lite-image	30 Jun 2026	Lite	Stable
Nano Banana 2.5 Flash	Not published	Not announced	Flash, by its name	Traces in Google Flow only

The Spicy-Mayo Tests Behind the Nano Banana 2.5 Rumours

The Flow trace is not the first sign of a new model. Testers have been using a candidate for two weeks without knowing its name.

What testers saw on Image Arena

Image Arena runs blind comparisons: a user writes a prompt, two anonymous models answer, and the user picks the better image. Labs use it to test models before launch. On 9 September 2026 several testers reported a new Google model there called spicy-mayo. LuminaBench wrote that it was “possibly Nano Banana 2.5 rather than just an update”.

Early reviews were positive but mixed. One tester, posting as synthwavedd, called it “a significant jump but not meaningfully better than GPT-Image 2.5, especially when it comes to world knowledge”. Another tester, Tim Jayas, ran the same prompt through Nano Banana 2 and spicy-mayo and reported that spicy-mayo “generated more accurate Minecraft UI, has better texture quality and was also faster”. These are individual tests, not benchmark scores.

Claims of partner access

From about 18 September, posts relayed by AI aggregator sites and Chinese tech outlets such as Mydrivers claimed that some partners could test Nano Banana 2.5 through Vertex AI. They said it offered minimal, medium and high thinking levels and image sizes up to 4K. None of the original posts shows documents, and Google has confirmed none of it. These details are unconfirmed and should not be relied on.

Why a medium thinking level would matter

The detail about thinking levels deserves attention because it can be checked against today’s model. Google’s developer documentation says Nano Banana 2 and Nano Banana 2 Lite support two levels, minimal and high, with minimal as the default. A middle level would let developers pay for extra reasoning on hard prompts, such as dense text or crowded scenes, without paying for the full high setting on every image. If Nano Banana 2.5 ships with three levels, that will be one of its most useful practical changes.

Why Google Flow Is the Natural Home for Nano Banana 2.5

Flow started as a video tool, but images now sit at the centre of how it works. That is why a trace there is plausible.

From filmmaking tool to creative studio

Google launched Flow at I/O on 20 May 2025 as an AI filmmaking tool built around its Veo video models. It replaced VideoFX and was first available to Google AI Pro and Ultra subscribers in the US. On 25 February 2026 Google redesigned it and moved the best parts of Whisk and ImageFX, its two image experiments, into Flow. Google said users had by then created more than 1.5 billion images and videos in Flow.

The redesign made image generation central. In Flow, users create an image, edit it with a lasso tool and plain instructions, and then use it as a frame or ingredient for a video. The Flow app’s Google Play listing still uses the package name Google gave Whisk, a small sign of that merger.

Nano Banana 2 is Flow’s free default

When Nano Banana 2 launched on 26 February, Google made it the default image model in Flow, “available to all Flow users for zero credits”. When Nano Banana 2 Lite launched in June, Google said it was coming to Flow as well. If a Nano Banana 2.5 Flash is coming, the Flow model picker is the obvious place to test it, and the free default slot is the obvious prize.

Agents, custom tools and Omni

On 19 May 2026 Google added Gemini Omni Flash, its model for generating and editing video, to Flow. It also added Flow Agent, custom Flow Tools built from plain-language instructions and mobile apps. Flow is now available in more than 140 countries. We covered the developer side of Omni when Google released Omni 1.1 Flash with upscaled 1080p and 4K output. In Flow, a still image often becomes the first frame of a clip, so a better image model improves the whole pipeline.

Date	Google Flow milestone	What it means for Nano Banana 2.5
20 May 2025	Flow launches for Veo, with Imagen for still images	Images start as a supporting feature
25 Feb 2026	Redesign merges Whisk and ImageFX into Flow	Images become central and free to generate
26 Feb 2026	Nano Banana 2 becomes the default at zero credits	Sets the slot a 2.5 model would take
19 May 2026	Omni Flash, Flow Agent, Flow Tools and mobile apps	Images feed agents and video models
30 Jun 2026	Nano Banana 2 Lite announced for Flow and other products	New image models are routed to Flow quickly
24 Sep 2026	Nano Banana 2.5 Flash traces reported in Flow	Testing appears to be under way

What Nano Banana 2.5 Has to Beat: Nano Banana 2 Today

Any successor will be judged against the model it replaces. Nano Banana 2 already covers most of what users asked for a year ago.

Resolution, references and consistency

According to Google, Nano Banana 2 keeps up to five characters and up to 14 objects consistent in one workflow. Its output ranges from 512 pixels to 4K, with a wide range of aspect ratios, including very tall formats such as 1:8. Google’s documentation says the Gemini 3 image models can combine up to 14 reference images in one request. Nano Banana 2 Lite gives up some of this: it produces 1K images only.

Grounding and reasoning

Nano Banana 2 can use Google Search while it works, and it is the only model in the family that also uses Google Image Search to find visual references. Nano Banana models also understand images as well as create them. The Gemini model underneath handles the computer vision side of an edit, such as locating the person to remove, before it draws the result. The Gemini 3 image models also produce unbilled “thought images” to plan a complex picture before the final render.

Known limits

Google’s documentation lists some limits. Image generation works best in 15 listed languages. The models do not accept audio input, and only the 3.1 Flash models accept video input. They “won’t always follow the exact number of image outputs” a user asks for. Search grounding cannot use real-world images of people from the web. Every image carries a SynthID watermark. A convincing Nano Banana 2.5 would need to improve text rendering, spatial logic and world knowledge, the areas where spicy-mayo testers still saw weaknesses.

Feature	Nano Banana 2 Lite	Nano Banana 2	Nano Banana Pro
Output resolution	1K only	512px, 1K, 2K, 4K	1K, 2K, 4K
Thinking levels	Minimal, high	Minimal, high	Reasoning built in
Google Search grounding	No	Web and Image Search	Web Search
Video input	Yes	Yes	No
Google’s positioning	Fastest, cheapest, about 4 seconds per image	Default all-rounder	Complex, professional work
Standard price, 1K image	$0.0336	$0.067	$0.134

Nano Banana 2.5 Pricing: What Current Rates Suggest

Google has published nothing about Nano Banana 2.5 pricing. The current price list does show the range a Flash-tier model is likely to fall into.

Today’s per-image prices

The Gemini API charges image output by tokens, and Google converts that into a price per image. At standard rates, Nano Banana 2 costs $0.045 for a 512-pixel image, $0.067 at 1K, $0.101 at 2K and $0.151 at 4K. Nano Banana 2 Lite costs $0.0336 per 1K image. Nano Banana Pro costs $0.134 per 1K or 2K image and $0.24 at 4K. The original Nano Banana costs $0.039 per image.

The Batch API halves those rates for jobs that do not need an instant answer. Nano Banana 2 drops to $0.034 at 1K and Nano Banana 2 Lite to $0.0168. Anyone forecasting a Nano Banana 2.5 budget should model batch and standard traffic separately, because the gap is a factor of two.

A thousand images at 1K

Multiplying the published per-image prices by 1,000 gives the cost of a thousand 1K images on each model.

Cost of 1,000 images at 1K resolution, standard Gemini API rates (September 2026)

Nano Banana Pro $134

Nano Banana 2 $67

Nano Banana, retiring 2 October $39

Nano Banana 2 Lite $33.60

Nano Banana 2 costs exactly half as much as Nano Banana Pro at 1K, and Nano Banana 2 Lite costs about 14% less than the original Nano Banana it replaces.

Price by resolution on Nano Banana 2

Resolution matters as much as the model. On Nano Banana 2, a 4K image costs more than three times a 512-pixel image.

Nano Banana 2 cost per 1,000 images by resolution, standard rates

4K (4096px) $151

2K (2048px) $101

1K (1024px) $67

0.5K (512px) $45

What a Nano Banana 2.5 price could look like

There is no published figure, and Google’s pricing does not follow a simple rule. Nano Banana 2 costs more per 1K image than the original Nano Banana ($0.067 against $0.039) but half the price of Nano Banana Pro, while Nano Banana 2 Lite undercuts both. A Nano Banana 2.5 Flash priced between Lite and Pro is the most likely outcome, but that is an inference from the tiers, not a leak. A Flow user would not see a per-image price at all, since Nano Banana 2 currently costs zero credits there. The key question for Flow is whether a 2.5 model keeps that free slot.

The 2 October Deadline That Will Not Wait for Nano Banana 2.5

For many developers, the most urgent date this month is a shutdown, not a launch.

What shuts down

Google’s pricing page carries a clear warning: Gemini 2.5 Flash Image, the original Nano Banana, “is deprecated and will be shut down on October 2, 2026”. Google tells developers to migrate to Gemini 3.1 Flash Image or Gemini 3.1 Flash Lite Image. That is eight days after the Flow sighting, and 402 days after the original model was announced.

Don’t wait for Nano Banana 2.5

It is tempting to skip a migration and wait for the next model. That would be a mistake. Google has not announced Nano Banana 2.5, and nobody outside Google knows its launch date, model ID or price. Calls to gemini-2.5-flash-image will stop working on 2 October whether or not a successor has arrived. Google names Nano Banana 2 Lite as the recommended replacement for the original model, and it is cheaper per image.

A short migration plan

The code change is usually small, because the request format is the same Gemini API. The testing takes more time. Run a sample of real prompts through both candidate models and compare text rendering, faces and product shots before switching. Check anything that depends on output size, since Nano Banana 2 Lite only produces 1K images and uses a different token count per image.

If you use the original Nano Banana for	Move to	Why
High-volume 1K images, thumbnails, drafts	Nano Banana 2 Lite	Google’s named replacement, $0.0336 per image
Marketing images needing 2K or 4K	Nano Banana 2	Supports 512px to 4K and Image Search grounding
Dense text, infographics, complex layouts	Nano Banana 2 at high thinking, or Nano Banana Pro	More reasoning for hard prompts
Batch jobs with no deadline	Either 3.1 model via the Batch API	Half the standard price
Waiting for Nano Banana 2.5	Migrate now, re-test later	The shutdown date is fixed; the launch is not

What Nano Banana 2.5 Could Change for Creative Teams

For people who use Flow rather than the API, the effect of a new model would show up in daily work.

Storyboards and keyframes

Flow’s workflow joins still images to motion. A team builds a mood board, generates keyframes with Nano Banana and turns them into clips with Veo or Omni. Each weak keyframe costs a video generation, which is where Flow’s credits go. A Nano Banana 2.5 that holds characters and props more steadily across frames would save regeneration time later in the process, even if each image arrives no faster.

Brand and product imagery

Product teams use Nano Banana to place real products in new scenes. The hard parts are keeping a logo correct, a label legible and a product the same shape between shots. The spicy-mayo reports praise texture and text in user-interface mock-ups. If Nano Banana 2.5 brings those gains to product shots, fewer images will need fixing in another editor, which is where much of the cost of AI imagery sits today.

Provenance stays attached

Every Nano Banana image carries an invisible SynthID watermark, and Google pairs SynthID with C2PA Content Credentials. In August Google began letting users remove the visible watermark from Gemini AI generations, but SynthID stays. Anyone can check an image with the tools described in our guide to verifying AI content with SynthID. Expect Nano Banana 2.5 images to carry the same marks.

What Nano Banana 2.5 Means for Businesses Building on Gemini

For most organisations, a trace in Flow changes nothing this week. It does suggest some sensible preparation.

If you use Gemini or Flow as a subscriber

There is nothing to do. When Google swaps the model behind the Gemini app or Flow, users get the new model automatically, as happened with Nano Banana 2 in February. Watch for changes to credit costs or defaults when a new model arrives, and keep copies of prompts that produce your house style so you can compare results.

If you build on the API

Pin model IDs in your code rather than relying on aliases, and keep a small test set of real prompts with the images you accepted. When Nano Banana 2.5 appears in the Gemini API, you can then measure it against your own work in an afternoon rather than relying on Arena anecdotes. Put the evaluation in your AI strategy as a standing task, because Google has released a new image model roughly every three to four months.

If you are choosing a supplier

Google is not the only option. Testers compared spicy-mayo with OpenAI’s GPT-Image 2.5, and the top image models change places often. Tools such as Google Pics, the Canva-style design tool built on Gemini and Nano Banana, place these models inside ordinary office workflows. Agencies and in-house marketing services teams should judge models by cost per usable image, not by the price per image generated.

Questions to ask before switching

Ask four questions. What does an accepted image cost once retries are counted? Does the model keep your products and faces consistent across a campaign? Which resolution does each use actually need? Can your legal team work with SynthID and Content Credentials on published images? These questions apply whether Nano Banana 2.5 arrives next week or next quarter.

When Could Nano Banana 2.5 Launch?

Nobody outside Google knows. The company’s recent pace and the stage of testing do give some clues.

Google’s release rhythm

Google has released a new Nano Banana model every three to four months. The chart below counts the days between each announcement, and the days from Nano Banana 2 Lite to the Flow sighting.

Days between Nano Banana announcements

Nano Banana 2 to Nano Banana 2 Lite 124 days

Nano Banana Pro to Nano Banana 2 98 days

Nano Banana to Nano Banana Pro 86 days

Nano Banana 2 Lite to the Flow traces (so far) 86 days

By 24 September, 86 days had passed since Nano Banana 2 Lite, the same gap as between the first two Nano Banana models. A launch in October would fit Google’s recent pace, but pace is not evidence. The first Nano Banana 2 Flash traces appeared 81 days before Nano Banana 2 launched, so traces can come well before a release.

Signals to watch

Four signals would show that a launch is close. The first is a model card or preview ID in the Gemini API model list. The second is a new entry in the Flow or Gemini app model picker. The third is Image Arena revealing spicy-mayo’s real name, as happens when anonymous models launch. The fourth is an announcement from Google DeepMind or on Google’s blog, The Keyword. Until one of these appears, Nano Banana 2.5 Flash remains a strong rumour.

What would count as confirmation

Confirmation means Google publishing a name, a model ID and a price. Screenshots, endpoint strings and Arena codenames all suggest that something is coming, but none of them says what it will be called or how it will perform. The Mayo episode showed that a leaked name can change before launch, so the final product could carry a different name from the one in the Flow code.

Nano Banana 2.5 FAQ

What is Nano Banana 2.5 Flash?

It is the name found in traces inside Google Flow, reported by TestingCatalog on 24 September 2026. By Google’s naming, it would be a Flash-tier successor to Nano Banana 2. Google has not announced it or published any specifications.

Is spicy-mayo the same model as Nano Banana 2.5?

Possibly. spicy-mayo is an anonymous Google image model that appeared on Image Arena on 9 September 2026. Testers link it to Nano Banana 2.5, and the codename fits Google’s condiment naming, but no one has confirmed the link.

When will Nano Banana 2.5 be released?

There is no announced date. The gaps between Google’s last four image models were 86, 98 and 124 days, and 86 days had passed since the last one when the Flow traces were reported.

Will Nano Banana 2.5 be free in Google Flow?

Unknown. Nano Banana 2 is Flow’s default image model and costs zero credits. Whether a Nano Banana 2.5 Flash would replace it on the same terms has not been said.

Should I wait for Nano Banana 2.5 instead of migrating from the original Nano Banana?

No. The original model, gemini-2.5-flash-image, shuts down on 2 October 2026. Google recommends Nano Banana 2 Lite or Nano Banana 2 as replacements. Migrate now and test Nano Banana 2.5 when it becomes available.