Nano Banana 2.5 Is Google’s Image Model — the “Spicy Mayo” Arena Test

Sofia Marenco

Sofia Marenco

Model Evaluation Lead

Published: September 10, 2026
Nano Banana 2.5 image model reference page

TLDRThe “Spicy Mayo” Arena appearance is the key clue: early testers report a jump over Nano Banana 2, but Google has not confirmed specs or release.

Nano Banana 2.5 is an unreleased Google-associated image-generation model first spotted in Image Arena under the codename “Spicy Mayo,” with its identity and specifications still unconfirmed. Early community reports describe a real improvement over Nano Banana 2, but say it does not clearly beat GPT-Image 2.5 on world knowledge. No confirmed price, public release date, API endpoint, context window, model size, or official benchmark is available.

Updated 2026-09-20: Third-party reports now claim partner Vertex access ahead of an expected public release next week (see the Update below).

Key Takeaways

  • Nano Banana 2.5 is an unconfirmed model name linked to Google through early Image Arena reports.
  • The model reportedly appeared under the codename Spicy Mayo during testing.
  • Early testers describe a Banana-2 Uplift, meaning an improvement over Nano Banana 2.
  • The same reports identify a World-Knowledge Gap against GPT-Image 2.5.
  • The current evidence consists of two dated public posts from two well-regarded community accounts, not a Google announcement.
  • Public access, pricing, licensing, technical specifications, and release timing are not yet confirmed.

What Is Nano Banana 2.5?

Nano Banana 2.5 is currently best understood as a suspected successor or update in Google’s Nano Banana image-model family. In the reporting available so far, the model has not been confirmed by an official Google or Google DeepMind announcement.

The strongest current clue is its appearance in Image Arena. On September 9, 2026, Md Ismail Šojal reported that Google had placed Nano Banana 2.5 in the Arena under the name Spicy Mayo. The post attributed the performance claims to testers and described a jump over Nano Banana 2. It also said the model remained behind GPT-Image 2.5 on world knowledge. The original report is available here

Google just slipped Nano Banana 2.5 onto Image Arena under the name spicy-mayo. - Testers say it’s a

Source: @0x0SojalSec

A second September 9, 2026 report from LuminaBench described Spicy Mayo as a new Google model in Arena. That report said the model was probably Nano Banana 2.5 rather than a routine update, while explicitly treating the identification as tentative. LuminaBench’s report is here

🚨A New Google Nana Banana model codenamed "Spicy Mayo" just appeared in Arena I’m pretty sure this

Source: @LuminaBench

Nano Banana 2.5 is an Arena-stage model report, not a confirmed public product. That distinction matters for developers deciding whether to design against it.

A separate September 5, 2026 report said Nano Banana 2.5 Flash was in development based on observations of the Gemini desktop app. That evidence is also a third-party leak report, not an official release notice. The relationship among Nano Banana 2.5, Nano Banana 2.5 Flash, and a possible Nano Banana 2.5 Lite remains unconfirmed.

Nano Banana 2.5 at a Glance

SpecificationCurrent status
DeveloperGoogle-associated, based on early reports; not officially confirmed
TypeImage-generation model, based on Image Arena identification; not officially confirmed
ModalityImage generation and possibly image editing; not yet confirmed
Reported codenameSpicy Mayo
Reported predecessor or baselineNano Banana 2
Context windowNot yet confirmed
Model sizeNot yet confirmed
PricingNot yet confirmed
AvailabilitySeen in Image Arena testing; no public release confirmed
API accessNot yet confirmed
LicenseNot yet confirmed
Official model cardNot yet confirmed
Numeric benchmark scoreNot yet confirmed

The table separates observations from specifications. A model appearing in an evaluation interface does not establish that developers can call it, redistribute it, or rely on its outputs in production.

How Nano Banana 2.5 Works / What Makes It Different

The internal mechanism of Nano Banana 2.5 is unknown. No architecture, parameter count, training description, input limit, output-resolution specification, or model identifier has been published so far.

What can be discussed is the current testing signal. The reports place the model in Image Arena, a comparative testing environment where users or testers can assess image models. The model was reportedly exposed under Spicy Mayo, rather than a confirmed public product name. That creates the first useful anchor: Spicy Mayo Track means the observed Arena identity, not a confirmed Google branding decision.

The second anchor is Banana-2 Uplift. Early testers reportedly saw a meaningful improvement over Nano Banana 2. However, the phrase describes community impressions rather than a measured score. The evidence gives no win rate, sample count, resolution, prompt set, or independent replication.

The third anchor is World-Knowledge Gap. Multiple summaries say the suspected model did not clearly surpass GPT-Image 2.5 on world knowledge. For an image model, this may refer to knowledge-dependent image prompts, factual visual composition, or related prompt-following behavior. The available reports do not define the test.

The available evidence supports a qualitative Banana-2 Uplift, but it does not support a numerical performance claim. Developers should treat the improvement as a test signal until examples and controlled comparisons appear.

The “same leak playbook as the original Nano Banana” comment suggests that Google may be allowing or tolerating limited evaluator exposure before a formal announcement. That is a pattern claim from a community post, not proof of Google’s release process.

What You Can Do With Nano Banana 2.5

At present, the only directly supported use is comparative testing inside Image Arena. The two posts discuss evaluation, not a public workflow, production integration, or a documented application.

If the model is later confirmed as a Nano Banana successor, likely evaluation areas include:

  • Image quality compared with Nano Banana 2.
  • Prompt adherence on factual or knowledge-heavy image requests.
  • Editing consistency across repeated prompts.
  • Text rendering inside posters, diagrams, and product images.
  • Response behavior across different image sizes and aspect ratios.

Those are test categories, not confirmed Nano Banana 2.5 capabilities. No verified public examples were included in the evidence bundle, and LuminaBench said examples were still forthcoming.

For engineering teams, the practical approach is to keep Nano Banana 2.5 out of production dependencies until an endpoint, model ID, usage policy, and repeatable evaluation set become available. A small internal test harness can still be prepared around representative prompts, provided every result is labeled as pre-release evidence.

How Nano Banana 2.5 Compares

GPT-Image 2.5 is the only direct competitor named in the early reports. The comparison remains provisional.

DimensionNano Banana 2.5GPT-Image 2.5
Current evidence statusEarly Arena reports; identity unconfirmedNamed comparison target in tester summaries
Relation to Nano Banana 2Reported improvementNot assessed in public reporting
World knowledgeReportedly behind GPT-Image 2.5Reportedly ahead of the suspected model
Numeric Arena scoreNot yet confirmedNot yet confirmed
Public Nano Banana 2.5 accessNot yet confirmedNot established by this page’s evidence

Based on early community testing, Nano Banana 2.5 may improve on Nano Banana 2 without overtaking GPT-Image 2.5 on world knowledge. That is the clearest current verdict, and it remains subject to verification.

The broader decision depends on the task. Teams focused on image editing consistency should wait for paired examples. Teams focused on factual visual prompts should specifically test the reported World-Knowledge Gap. A separate GPT Image 2.5 versus Nano Banana 2 analysis provides context for the older family comparison, but it does not verify Nano Banana 2.5.

Availability: How to Access Nano Banana 2.5

Nano Banana 2.5 does not have confirmed public access in the sources we reviewed. The model’s appearance in Image Arena should not be treated as general availability.

Access routeCurrent status
Image ArenaReported testing route under Spicy Mayo
Google Gemini appNano Banana 2.5 access not yet confirmed
Google AI StudioNano Banana 2.5 access not yet confirmed
Google developer APINano Banana 2.5 endpoint and model ID not yet confirmed
kie.ai APIkie.ai offers access to related Nano Banana models, including Nano Banana 2; Nano Banana 2.5-specific hosting is not confirmed
Downloadable weightsNot yet confirmed
Open-source licenseNot yet confirmed

For developers who need a currently listed Nano Banana-family route, kie.ai’s Nano Banana 2 API page is the relevant available option in the supplied catalog. It should not be presented as proof that the unreleased 2.5 variant is available there.

The official Google app, AI Studio, and developer documentation remain the routes to monitor for a first-party announcement. No official resource in the evidence bundle confirms a Nano Banana 2.5 model string.

What We Don’t Know Yet

Several important questions remain open as of September 10, 2026:

  • Whether Spicy Mayo is actually Nano Banana 2.5 or a separate internal update.
  • Whether Nano Banana 2.5 and Nano Banana 2.5 Flash are the same model.
  • Whether a Nano Banana 2.5 Lite variant exists.
  • Whether the model generates images only or also supports image editing.
  • What context window, image-input limit, output resolution, and latency it supports.
  • Whether the reported Banana-2 Uplift holds across a controlled prompt set.
  • Whether the World-Knowledge Gap reflects factual reasoning, prompt adherence, or another Arena category.
  • When Google might announce the model.
  • Whether developers will receive API access, consumer-app access, or both.
  • How the model will be priced and licensed.

The evidence base is especially small. It rests on two public posts from two named community accounts, and neither gives a numeric score. That is enough to justify monitoring the term, but not enough to publish a technical specification sheet as fact.

Frequently Asked Questions

What is Nano Banana 2.5?

Nano Banana 2.5 is an unreleased Google-associated image model reportedly seen in Image Arena under the codename “Spicy Mayo.” Its identity, specifications, price, and public release status are not officially confirmed.

Is Nano Banana 2.5 released?

Nano Banana 2.5 is not confirmed as publicly released. Early reports describe a model being tested in Image Arena, but they do not establish access through Google’s consumer apps, developer API, or an official model page.

What is the Spicy Mayo Nano Banana model?

Spicy Mayo is the codename attached to a new Google-associated model seen in Image Arena. Two early reports tentatively connect it with Nano Banana 2.5, but Google has not confirmed that identity.

How does Nano Banana 2.5 compare with Nano Banana 2?

Early community reports describe Nano Banana 2.5 as a meaningful improvement over Nano Banana 2. The reports provide no numeric benchmark, Arena score, or verified capability breakdown.

Is Nano Banana 2.5 better than GPT-Image 2.5?

Nano Banana 2.5 is not established as better than GPT-Image 2.5. Early tester summaries say it may improve on Nano Banana 2 while still trailing GPT-Image 2.5 on world knowledge.

How much does Nano Banana 2.5 cost?

Nano Banana 2.5 has no confirmed public price. The supplied reports do not publish per-image pricing, subscription pricing, token rates, or an official API billing schedule.

What to watch next: a Google announcement, a confirmed model identifier, and reproducible Image Arena examples should resolve the model’s identity and performance. Pricing, API availability, and a public model card will determine whether Nano Banana 2.5 is useful for production systems.

Update — 2026-09-20

On September 18, @lyraxana and @pankajkumar_dev claimed that partners had access through Vertex under the model identifier nano-banana-2.5. Their reports described minimal, medium, and high thinking levels, with 512, 1K, 2K, and 4K image sizes, and expected a broader release during the week of September 21. @Ananth7e separately repeated the “next week” expectation. These remain third-party reports; Google has not officially confirmed the identifier, specifications, release date, or broad public availability.

The newer task-specific tests are mixed. In a same-prompt Minecraft comparison, @TimJayas reported better texture quality, speed, and UI accuracy than Nano Banana 2; against GPT-Image 2.5, the same tester gave GPT-Image 2.5 the UI-accuracy win while finding Nano Banana 2.5 more visually appealing (comparison). @LuminaBench later described the Arena version as strong at text but weak on logical tasks, following a reported Rubik’s-cube failure. These observations are not standardized benchmark results.

Try related models on Kie.ai: Nano Banana 2 Lite, Nano Banana 2 and Nano Banana 2.1.

Sofia Marenco

About Sofia Marenco

Sofia stress-tests new models on coding and reasoning benchmarks and reports what holds up.

View all posts by Sofia Marenco