
Editorial concept: following an anonymous image-model signal from early tests to a possible named release. This is not model output.
Nano Banana 2.5 is the name now attached to spicy-mayo, an image model first discussed as an anonymous Image Arena candidate. The story moved forward on September 20, 2026, when one source reported that partners had received a model named nano-banana-2.5 through Vertex, with three thinking levels and output sizes from 512 to 4K. Google has not added that name to its public Nano Banana lineup, but the newer report is more specific than the original Arena discussion and makes this a release signal worth watching.
The useful takeaway: Nano Banana 2.5 may be a refinement positioned between rapid everyday generation and more deliberate high-quality work. The reported controls suggest that users could choose how much reasoning and resolution a job needs instead of switching models for every task.
Nano Banana 2.5 key takeaways
- Spicy-mayo is the leading codename. It appeared in early September discussion around an anonymous Image Arena model.
- The latest report names a real distribution path. It says Vertex partners received access under
nano-banana-2.5. - Three thinking levels were reported:
minimal,medium, andhigh. - Four output sizes were reported: 512, 1K, 2K, and 4K.
- The public Google model catalog still stops at Nano Banana 2 and Nano Banana 2 Lite. There is no public Nano Banana 2.5 listing there as of September 21.
- BananaAI does not currently offer a verified Nano Banana 2.5 model. This guide will track the model as the release picture develops; use Nano Banana 2 for an available workflow today.
Why Nano Banana 2.5 is suddenly news
The first serious signal arrived on September 9. An AI commentator using the name synthwavedd said Google DeepMind was testing a model called Nano Banana 2.5, codenamed spicy-mayo, in Image Arena. A second post the same day described the model as real but said its exact version was unclear: it might be 2.5, a Flash Lite model, or something else.
The September 20 report added the details that were missing from that first wave. Lyra's original post claimed that spicy-mayo had been deployed to partners through Vertex under the explicit name nano-banana-2.5. It also listed thinking levels and image sizes, and said a release was planned for the following week. Later posts from Mark Kretschmann and Kol Tregaskes amplified that same source rather than independently confirming it.
| Date | Signal | What changed |
|---|---|---|
| Sep 9, 2026 | Spicy-mayo discussed as an Image Arena model | A test candidate appears, but the public name is uncertain |
| Sep 20, 2026 | Partner access reported through Vertex | The claim gains a specific model name, distribution channel, controls, and size tiers |
| Sep 21, 2026 | No matching entry in Google's public model catalog | Public access, final branding, pricing, and launch timing remain open |
This is why the current story is stronger than a random version-number guess. It now contains a codename, a claimed internal model ID, a plausible partner channel, and a coherent set of user-facing controls. It is still one original report, however; retweets and quote posts increase reach, not the number of independent sources.
What the reported controls could change for creators
If the reported configuration reaches a public product, its most interesting feature may not be a single quality jump. It may be the ability to choose a different balance of reasoning, speed, and output size for each stage of a creative job.
Minimal thinking for drafts and exploration
A minimal mode would make sense for thumbnailing, visual brainstorming, background variations, and prompt exploration. At this stage, creators usually value fast feedback more than perfect typography or tiny details. A 512 or 1K output could reduce the cost of ideas that may be discarded after one look.
Medium thinking for routine production
A middle setting could become the practical default for product concepts, social graphics, character scenes, and normal image edits. The useful question is whether medium reasoning preserves subject identity and follows multi-part instructions reliably without the wait associated with a maximum-quality mode.
High thinking and 4K for final assets
A high setting paired with 2K or 4K output points toward polished posters, product campaigns, diagrams, presentation graphics, and other assets where text, layout, and small objects matter. It could also make the model easier to fit into a draft-to-final workflow: explore quickly, then spend more computation only on the selected direction.

Conceptual workflow: the reported controls could let creators spend less time on drafts and more reasoning on a selected final. This illustration was not generated by Nano Banana 2.5.
The real test will be whether these levels create predictable tradeoffs. A useful control should make high visibly better on difficult prompts, not merely slower. It should also keep edits compatible across stages so a draft can move to a final render without changing the composition or subject unexpectedly.
How Nano Banana 2.5 may fit Google's current lineup
Google's Gemini image generation guide currently lists four public Nano Banana models. The official Gemini Image family page presents the same active family without a 2.5 entry.
| Public name | Technical name | Current role |
|---|---|---|
| Nano Banana | Gemini 2.5 Flash Image | The original fast image model, now treated as the legacy option |
| Nano Banana Pro | Gemini 3 Pro Image | Premium precision, knowledge, localization, and production control |
| Nano Banana 2 | Gemini 3.1 Flash Image | The versatile generalist balancing speed, 4K output, knowledge, and editing |
| Nano Banana 2 Lite | Gemini 3.1 Flash-Lite Image | The fastest, lowest-cost option for scale and rapid generation |
That lineup leaves room for several interpretations. Nano Banana 2.5 could be a stronger successor to Nano Banana 2, a quality tier that borrows more behavior from Pro, or simply the working name of a candidate whose public brand will change. The reported thinking levels make the first two possibilities interesting because they would let one model stretch across draft and final workflows.
The name also creates an easy source of confusion. Nano Banana 2.5 is not the same phrase as Gemini 2.5 Flash Image. Google launched Gemini 2.5 Flash Image as the original Nano Banana in 2025; its model card still uses gemini-2.5-flash-image. A new product called Nano Banana 2.5 would be a later version in the consumer-facing Nano Banana sequence, not a relaunch of that older technical model name.
Are Nano Banana 2.5, spicy-mayo, and 2.5 Flash the same model?
For now, spicy-mayo is the codename most directly connected to the Nano Banana 2.5 reports. The phrase Nano Banana 2.5 Flash has also circulated, but the latest source used nano-banana-2.5, not nano-banana-2.5-flash. There is not enough separation to treat 2.5 Flash as a second product.
The cleanest interpretation is:
- Spicy-mayo is the test codename.
- Nano Banana 2.5 is the reported partner-facing model name.
- Nano Banana 2.5 Flash is an informal variant people may be using to describe its expected speed or Gemini family position.
If Google later publishes separate 2.5 and 2.5 Flash models, this guide will split their roles. Until then, treating all three phrases as one developing story avoids inventing a model family from naming speculation.
Can you use Nano Banana 2.5 now?
There is no public Nano Banana 2.5 choice in Google's current Gemini Image catalog, and BananaAI has not connected a verified provider model under that name. The reported Vertex access appears to concern selected partners rather than a general consumer rollout.
If you need to create or edit images now, Nano Banana 2 is the closest current option. Google describes it as the all-around Gemini image model, with 4K generation, web-informed knowledge, text rendering, multiple reference images, and conversational editing. This guide will be updated with a verified generator and model-specific examples when direct Nano Banana 2.5 access becomes available.
What would confirm the Nano Banana 2.5 release?
The next meaningful evidence will come from one of five places:
- A Google or Google DeepMind announcement using the Nano Banana 2.5 name.
- A new model ID in the Gemini API documentation or Google AI Studio.
- A public Vertex AI catalog entry with access and regional details.
- A reveal connecting spicy-mayo's Arena identity to a Google model.
- Public pricing, limits, or product availability in Gemini, Search, or another Google surface.
Those events would answer more than the launch-date question. They would establish whether thinking levels are exposed to ordinary users, whether 512-to-4K choices apply to both generation and editing, and whether 2.5 replaces Nano Banana 2 or occupies a separate quality tier.
Nano Banana 2.5 FAQ
What is Nano Banana 2.5?
Nano Banana 2.5 is the reported name of a new Google image model associated with the spicy-mayo test codename. The latest report says partners received Vertex access with three thinking levels and four image sizes. It does not yet appear in Google's public model lineup.
When will Nano Banana 2.5 be released?
The September 20 source said it was planned for the following week, which points to late September 2026. Google has not published a launch date, so that window should be treated as a near-term report rather than a scheduled event.
Will Nano Banana 2.5 support 4K images?
The reported partner configuration includes 512, 1K, 2K, and 4K image sizes. Public documentation will need to show which sizes are available in consumer products and whether editing supports the same options.
What are Nano Banana 2.5 thinking levels?
The leak lists minimal, medium, and high. They likely control how much reasoning the model applies before producing an image, creating a tradeoff between fast drafts and more deliberate final output. Their exact behavior has not been documented publicly.
Is Nano Banana 2.5 free?
No public pricing or free allowance has been announced for this model name. Availability could differ between Gemini, Vertex, Google AI Studio, and third-party products if it launches across several channels.
Is Nano Banana 2.5 better than Nano Banana 2?
There is no same-prompt public comparison that establishes that yet. The reported thinking controls suggest a broader quality-versus-speed range, but the most important tests will be instruction following, text accuracy, subject consistency, editing stability, latency, and cost on the same tasks.
Update log
- September 21, 2026: Added the reported Vertex partner deployment,
nano-banana-2.5name, thinking levels, 512-to-4K sizes, source relationship, and current official model-family comparison. - September 10, 2026: Initial spicy-mayo and Image Arena signals reviewed.
