Nano Banana 2.5 Tool? spicy-mayo Access on LMArena

Spicy-mayo is attracting attention because community observers say the name has appeared during anonymous image battles on LMArena. A separate leak report points to the phrase Nano Banana 2.5 Flash in Google-related code. Those are two interesting clues, but they are not yet one confirmed story.

As of September 10, 2026, Google has not announced a model called Nano Banana 2.5 Flash, and Arena has not published a provider reveal for spicy-mayo. The current public text-to-image leaderboard does not list either name. The responsible answer is therefore: spicy-mayo may be a Google image checkpoint, but the evidence does not yet prove that it is Nano Banana 2.5 Flash.

This page tracks what is known, what remains missing, and how to test an anonymous model without letting a single impressive image decide the verdict. Nano Banana 2.5 is not confirmed or available on Image3. GPT Image 2 is a currently available alternative—not Nano Banana 2.5 or spicy-mayo.

Can you use a Nano Banana 2.5 tool now?

No verified public Nano Banana 2.5 tool is available under that exact name as of September 10, 2026.

If your goal is to generate or edit images today, compare the current GPT Image 2 model workflow or open the preselected Image3 Workspace. These links do not provide access to Nano Banana 2.5; they provide a clearly labeled alternative while the rumored model remains unconfirmed.

Last checked: September 10, 2026.

Quick answer: is spicy-mayo Nano Banana 2.5 Flash?

Not on the available evidence.

The strongest case for the connection is circumstantial:

A product-level identification needs a bridge between the two names. That bridge could be an Arena provider reveal, a Google release note, a public API identifier, a model card, or a statement from either organization. None was available at the time of this review.

What has actually been verified

Three facts have primary or directly inspectable support.

Sources checked for this update: Google's image-generation documentation, Arena's original nano-banana reveal, Arena's public text-to-image leaderboard, and dated community reports. The source URLs, access dates, and usage-rights records are retained in the article's research ledger.

First, nano-banana was a real LMArena testing codename. Arena's own retrospective says the alias used during pre-release testing was later revealed as Google's Gemini 2.5 Flash Image. That history explains why a food-themed anonymous name can draw serious attention.

Second, Google's current public naming has already moved beyond the original Nano Banana. Its image-generation documentation lists four products:

• Nano Banana 2 Lite, mapped to gemini-3.1-flash-lite-image
• Nano Banana 2, mapped to gemini-3.1-flash-image
• Nano Banana Pro, mapped to gemini-3-pro-image
• Nano Banana, the legacy gemini-2.5-flash-image

Third, the public Arena text-to-image leaderboard did not contain spicy-mayo, instant-ramen-v2, or Nano Banana 2.5 Flash when checked on September 10. That absence does not disprove private or rotating battle tests. It only means there is no stable public leaderboard row that readers can use as attribution evidence today.

Where the Nano Banana 2.5 Flash rumor came from

The Nano Banana 2.5 Flash phrase did not start with an official Google launch page. It appeared in a September 2026 report about code traces found while Google was testing a broader Gemini desktop experience. The report showed the phrase and described the work as closed testing with no clear public-release timeline.

That makes the string useful evidence of an internal or experimental label, but code strings are fragile clues. They can describe a feature flag, a UI placeholder, a compatibility target, an abandoned name, or a model that ships under a different brand. They do not establish price, API access, quality, release date, or architecture.

The most accurate wording is:

Reported: a Google-related code trace included "Nano Banana 2.5 Flash."
Not reported: a public model card, API ID, launch date, price, or capability table.

This distinction matters because a search-friendly name can harden into a false fact once enough sites repeat it without carrying the original uncertainty.

What spicy-mayo appears to be

Spicy-mayo appears to be an anonymous or pre-release image checkpoint label used in Arena testing, based on posts from people who follow rotating Arena models. One account described it as a new Nano Banana model or update. Another said an earlier label, instant-ramen-v2, had been renamed to spicy-mayo and characterized it as a Nano Banana 2 Lite update.

Those accounts are useful discovery signals, not provider confirmation. They can tell us what a tester saw on a particular day. They cannot independently prove the lab behind the model, whether two aliases point to the same weights, or whether a checkpoint will ship.

The defensible one-sentence description is:

Spicy-mayo is a community-reported LMArena image-test alias whose provider and final product name remain unconfirmed.

That sentence is less exciting than “Google just launched Nano Banana 2.5,” but it will age better and is less likely to mislead readers arriving from search.

Why people suspect Google

The Google theory rests on pattern matching.

Arena has already hosted Google's original Gemini 2.5 Flash Image under the nano-banana codename. The new alias is another food name. Community trackers also connect it to instant-ramen-v2, while a separate code report mentions Nano Banana 2.5 Flash at almost the same time.

The timing and naming pattern make the theory reasonable. They do not make it verified. Anonymous evaluation platforms deliberately hide model identity so people judge outputs rather than brands. A playful alias can be chosen by the platform, the provider, or a testing system. Similar naming is not a cryptographic signature.

For this reason, the evidence should be graded rather than flattened:

The naming trap: 2.5 does not mean Gemini 2.5

This rumor is unusually easy to misunderstand because Google's official product names already contain both “Nano Banana 2” and “Gemini 2.5 Flash Image.”

The original Nano Banana maps to Gemini 2.5 Flash Image. Nano Banana 2 maps to Gemini 3.1 Flash Image. Therefore, the rumored phrase Nano Banana 2.5 Flash cannot safely be interpreted as another spelling of Gemini 2.5 Flash Image.

Keep the labels separate:

Official legacy model: Nano Banana = gemini-2.5-flash-image
Official current model: Nano Banana 2 = gemini-3.1-flash-image
Rumored label: Nano Banana 2.5 Flash = no verified public API ID
Arena alias: spicy-mayo = no verified provider

This taxonomy is not cosmetic. If a developer copies an invented model string into an API request, the request fails. If a buyer assumes a rumor page describes an available product, the page creates the wrong expectation. Search coverage should help readers resolve the ambiguity, not amplify it.

What the current evidence cannot tell us

No reliable source currently establishes any of the following for spicy-mayo or Nano Banana 2.5 Flash:

• Public availability in Gemini, AI Studio, Vertex AI, or an API
• A stable model ID
• Input-image limits or supported reference types
• Maximum output resolution
• Text-rendering accuracy
• Multi-turn editing behavior
• Latency, rate limits, or price
• SynthID behavior or other provenance details
• Safety policy or commercial-use terms
• A leaderboard score based on a stable sample

Any article publishing those details today would either be confusing an existing official Nano Banana model with the rumored label or extrapolating from a few anonymous outputs.

This is also why screenshots alone are weak proof. A screenshot can show that an alias was displayed. It cannot show the hidden provider, model build, system prompt, sampling configuration, or whether the label later changed.

How to test spicy-mayo fairly on Image Arena

If spicy-mayo appears in an anonymous battle, the goal is not to guess the logo behind it. The goal is to create a record that remains useful after the reveal.

Use five task families: typography, instruction following, reference preservation, local editing, and multi-turn consistency. Run at least three samples for each family. Score the outputs before the model labels are revealed. Keep prompt wording, source images, and scoring rules fixed.

For every battle, record:

Date and local time:
Arena mode:
Prompt version:
Input-image hashes or filenames:
Left output score before reveal:
Right output score before reveal:
Revealed labels:
Visible errors:
Export or screenshot filename:

Do not combine “looks better” and “followed the brief” into one score. A dramatic image may win on first impression while changing a product, misspelling a headline, or ignoring a requested edit.

Five prompts that expose real differences

The following suite is designed to test control, not just visual taste. Use the same wording and inputs across models.

1. Typography and layout

Create a 4:5 editorial poster for a fictional astronomy museum.
Use exactly this headline: NIGHT MAPS
Use exactly this subheading: One sky, four seasons
Place the headline in the upper third and the subheading directly below it.
Show a precise circular star chart in the lower half.
No other words, letters, logos, or numbers.

Score exact spelling, extra text, hierarchy, and whether the star chart remains legible.

2. Dense instruction following

Create a square product photograph of one matte-red travel mug on a pale gray table.
The mug must have a black lid, one silver button, and no logo.
Place a folded navy map behind the mug and a brass compass to its right.
Use soft window light from the left.
Do not add people, plants, text, steam, extra cups, or reflections of the camera.

Count instruction passes and failures. Do not award a pass because the overall photograph is attractive.

3. Reference preservation

Use the attached product image as the source of truth.
Keep its silhouette, cap, label proportions, material, and visible color unchanged.
Place the same product on a wet black stone beside a mountain stream at dawn.
Match the new lighting and reflections to the environment.
Do not redesign the product or invent label text.

Inspect geometry and small identity details, not only color similarity.

4. Local edit discipline

Change only the background wall from white to deep forest green.
Keep the person, face, hair, pose, clothes, chair, floor, crop, shadows, and camera angle unchanged.
Do not retouch skin or alter the lighting on the subject.

Use a pixel or overlay comparison when possible. Local-edit quality is the amount of requested change divided by unintended change.

5. Multi-turn consistency

Turn 1: Create a three-quarter view of a small orange delivery robot with two black wheels, one square camera, and a white cargo box.
Turn 2: Show the same robot from the left side. Preserve every design detail.
Turn 3: Place the same robot in light rain at night. Change only the environment and lighting.
Turn 4: Add the number 27 to the cargo box. Change nothing else.

Score identity drift after each turn. A model that produces a strong first frame but redesigns the robot on turn three is not consistent.

Use a scorecard, not a vibe

A simple 100-point rubric makes blind comparisons more useful:

Store raw scores as well as the total. If a model wins 88 to 84, readers should be able to see whether the difference came from prettier lighting or materially better constraint following.

The clip below shows a currently available Image3 GPT Image 2 workflow. It is not a spicy-mayo or Nano Banana 2.5 output, and it is not evidence about either rumored name. Its purpose is to demonstrate a repeatable way to inspect consistency across several controlled frames.

What would confirm the model's identity

The rumor moves from plausible to confirmed only when a primary source supplies a durable identifier. Watch for one or more of these:

1. Arena publishes a post-reveal note that maps spicy-mayo to a provider and official model.
2. Google adds Nano Banana 2.5 Flash to a release note or developer model list.
3. A public API exposes a documented model ID with matching capabilities.
4. The Gemini app or AI Studio displays the name in a stable selector with product documentation.
5. A model card describes evaluation results, limitations, safety behavior, and provenance.

A leaderboard row would help with performance context but would not, by itself, prove that spicy-mayo and Nano Banana 2.5 Flash are the same. The name mapping still needs to come from Arena or Google.

Should you wait for Nano Banana 2.5?

Not for production work.

Rumored checkpoints can disappear, change names, or ship with different limits. Teams should evaluate available models against the job they need to complete today. For example, compare typography accuracy, reference-image preservation, edit isolation, latency, and cost on your own prompt set.

Image3 is an independent image workflow site and is not affiliated with Google, Gemini, or LMArena. Its current generation path focuses on GPT Image 2 workflows. You can review what GPT Image 2 is, see how mixed-input media systems are changing in the Gemini Omni explainer, or browse copy-ready structures in the GPT Image 2 prompt library.

That comparison also guards against a common launch-cycle mistake: treating a future model as automatically better at your specific task. A new checkpoint might improve speed while trading away reference capacity, or improve general aesthetics without fixing exact text. Only the workload decides whether an upgrade matters.

When this page will be updated

This page will change when a durable source resolves one of the open questions. The clearest update triggers are an Arena provider reveal, a Google release note or model card, a documented public API ID, or a stable product selector in Gemini or AI Studio.

Until one of those records appears, the title keeps its question mark and the availability table stays marked unverified. A new screenshot, repost, or food-themed alias alone is not enough to change the conclusion.

Bottom line

Spicy-mayo is worth watching. The original nano-banana story shows that anonymous Arena testing can precede an important Google image release. A current code report also makes Nano Banana 2.5 Flash a traceable rumor rather than a phrase invented after the fact.

But the decisive link is still missing. No official source currently maps spicy-mayo to Google, Nano Banana 2.5 Flash, Nano Banana 2 Lite, or any other final product. Treat output comparisons as model testing, not identity proof.

While that evidence develops, use a fixed scorecard and repeatable prompts. If you need to produce images now, return to Image3, open the GPT Image 2 image generator, or start a controlled test in the Image3 Workspace. Those are current alternatives, not a Nano Banana 2.5 tool.