WaveSpeedAI
← Blog/News & Insights

Nano Banana 3 Pro Spotted: What the Gempix 3 Pro Leak Tells Us

A 'Gempix 3 Pro' label has surfaced in Google's image citation system, pointing to a Nano Banana 3 Pro built on Gemini 4 Argon. Here is what is reported, what is speculation, and what builders should do now.

By WaveSpeedAI5 min read
#nano-banana#google#gemini-4#image-generation#leak

Google’s next image model may already be in the works. On October 12, 2026, TestingCatalog reported that a “Gempix 3 Pro” label appeared in Google’s image citation functionality. Going by Google’s naming pattern, that points to Nano Banana 3 Pro.

Google has confirmed nothing. Here is a clear split between what was found and what is guesswork.

What was found

  • A new internal label. “Gempix” is an internal name Google has used before. TestingCatalog notes that Gempix 2 was the label for the previous Nano Banana generation, so Gempix 3 Pro most likely maps to the next one.
  • It is wired into the image citation system. Gempix 3 Pro has been added to Google’s existing image citation functionality. That system checks whether a generated image closely matches an existing reference image, such as content from Shutterstock or Pinterest, and can block the generation when it does.
  • A link to Gemini 4 Argon. The findings indicate the new image model will be based on Gemini 4 Argon, the next Gemini generation. Google AI Ultra subscribers are expected to get Gemini 4 Argon after its trusted-testing phase.

What is speculation

  • The product name. “Nano Banana 3 Pro” is an inference from the naming convention. The article says Gempix 3 Pro “could be” it.
  • Quality gains. Stronger image generation and editing is the expected result of a new base model. No samples or benchmarks have been published.
  • Timing. TestingCatalog calls November to December a plausible window. Google has announced no date.
  • Ultra-first access. The Ultra expectation is tied to Gemini 4 Argon, not explicitly to the image model.

Why the citation detail matters

The citation hook is the most useful signal in the report. A model that is only an experiment rarely gets connected to content-protection systems. Wiring Gempix 3 Pro into the reference-matching check suggests it has moved past a lab prototype and is being prepared for real users.

It also hints at product behavior. If a prompt produces an image too close to a known reference, a request may be blocked instead of returned. Expect this to affect prompts that imitate stock photography or popular pinned imagery.

What it could mean for the Nano Banana line

The pattern would match how Google has shipped image models so far: each Gemini generation gets a matching Nano Banana release. The earlier Nano Banana models sat on Gemini 3 generation bases, so a Gemini 4 base is the natural next step.

If that holds, the areas worth watching are the ones where current models still slip:

  • Text rendering in dense layouts, posters and infographics
  • Multi-image fusion, where several references must stay consistent
  • Editing precision, changing one element without disturbing the rest

None of this is confirmed. Treat it as a checklist for when real samples appear, not as a prediction of results.

What to do now

  1. Don’t plan around a date. A November to December window is a guess.
  2. Keep your integration model-agnostic. If your product calls an image model, keep the model ID in configuration so you can swap in a new one and compare outputs side by side.
  3. Build a small test set today. Collect 20 to 30 prompts that fail or look weak on your current model: text-heavy designs, multi-reference edits, product shots. Re-run them the day a new model is available.
  4. Use the current generation in the meantime. Our Nano Banana Pro complete guide and Nano Banana 2 API guide cover what the existing models do well. For the base model behind the rumor, see our Gemini 4 status tracker.

Try Nano Banana on WaveSpeedAI today

You do not have to wait for the next generation to build and test your pipeline. Two Nano Banana models are available through the WaveSpeedAI API now:

ModelModel pathPrice per image
Nano Banana 2.1google/nano-banana-2.1/text-to-image$0.04 (1K), $0.06 (2K), $0.135 (4K)
Nano Banana Progoogle/nano-banana-pro/text-to-image$0.14 (1K), $0.14 (2K), $0.24 (4K)
  • Nano Banana 2.1 is the fast, low-cost option. It supports 14 aspect ratios, optional web search and image search grounding, and PNG or JPEG output. Image editing with up to 14 reference images is available at google/nano-banana-2.1/edit, at the same $0.04 / $0.06 / $0.135 per image plus $0.002 for each reference image.
  • Nano Banana Pro is the quality-first option for text-heavy layouts and detailed edits. google/nano-banana-pro/edit also accepts up to 14 reference images.

Both use the same request pattern: send a POST to https://api.wavespeed.ai/api/v3/<model-path> with your API key, then poll https://api.wavespeed.ai/api/v3/predictions/{id}/result until the status is completed.

curl -X POST "https://api.wavespeed.ai/api/v3/google/nano-banana-2.1/text-to-image" \
  -H "Authorization: Bearer $WAVESPEED_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"prompt": "A product poster with the headline FRESH DROP in bold type", "resolution": "2k"}'

Because the model path is just a string, moving to a new Nano Banana release later means changing one value and re-running your test set. Prices above come from the live model pages on the date of this post and can change; the page shows the current estimate before you generate.

Bottom line

A “Gempix 3 Pro” entry in Google’s citation system is real evidence that a next-generation image model is being prepared. The Nano Banana 3 Pro name, the Gemini 4 Argon base and the timeline are still reports and inference. We will update this post when Google says something official.

Share