Nano Banana Pro is Google’s most advanced image-generation and editing model, built on Gemini 3 Pro. It extends the original Nano Banana with significantly improved multimodal reasoning, real-world grounding, and...
This model is a preview release from Google of an upcoming Gemini 3 Pro variant focused on image understanding. It accepts image and text inputs and generates text outputs. The "Nano Banana Pro"…
The model can describe image content in detail, answer questions about objects, scenes, colors, and text visible in the image (e.g., reading signs or labels). It supports visual reasoning tasks such as comparing two images, identifying differences, or extrapolating from visual cues. It can also analyze diagrams and charts, though its accuracy on complex scientific figures may vary.
Strong use cases include: 1) Real‑time image captioning for accessibility tools; 2) Visual search and product classification in e‑commerce; 3) Document processing (forms, receipts, invoices) with handwritten or typed text; 4) Educational tools that explain diagrams or photographs. The model performs best with clear, well‑lit images and specific, granular prompts.
If your task does not involve images (e.g., pure text generation, summarization, translation), a text‑only model (like standard Gemini or Llama variants) will be more cost‑effective. For simple image classification that does not require natural language reasoning, a dedicated vision model with lower latency may be better. Also, if you need lower latency for high‑volume requests, a smaller multimodal model might be preferable.
As a preview, function calling support is not confirmed for this model. OrcaRouter's API supports the same tool definitions as OpenAI, but the underlying model may not reliably execute function calls. Test thoroughly before relying on tool use. Structured output (JSON mode) is supported via the OpenAI‑compatible format, but output quality varies.
Benchmark scores for Nano Banana Pro (Gemini 3 Pro Image Preview) have not been publicly released. As a preview model, it may not be evaluated on standard benchmarks like MMLU, VQAv2, or COCO Captions. Developers should run their own evaluation on representative data to assess performance. Expect improvements in the final Gemini 3 Pro release.
Latency depends on image size, input length, and OrcaRouter's current load. Image processing adds overhead compared to text‑only models. On average, a request with one medium‑resolution image and 100 text tokens may take several seconds for the first token and then stream the rest. There is no published tokens‑per‑second figure for this preview. Use smaller images and batch requests to minimize latency.
The model excels at identifying objects, people, and text within images. It can reason about spatial relationships and answer questions that require combining visual and textual information. Early feedback indicates good performance on photo‑based queries and document understanding. Its large context window allows multi‑image conversations.
As a preview, the model may produce unexpected outputs or hallucinate details about images (e.g., claiming objects that are not present). It may struggle with low‑resolution, blurry, or highly abstract images. Complex multi‑step visual reasoning (e.g., math equations from handwriting) can be unreliable. The model does not support audio or video input. There is no fine‑tuning option for this preview.
Pricing is set by OrcaRouter based on per‑token costs for the google provider. Input tokens are typically cheaper than output tokens. Image tokens consume significantly more input tokens than text—each image is tiled and processed. Consult OrcaRouter's official pricing page for current rates. There is no free tier for this model; you pay per request.
Because image processing is token‑intensive, costs can accumulate quickly if you send many high‑resolution images. To manage costs: reduce image resolution, limit the number of images per request, and use short prompt text. For tasks where images are not essential, consider a text‑only model. OrcaRouter may offer caching for repeated image embeddings (check documentation for details).
OrcaRouter may implement caching for frequently used image embeddings, but this preview model's caching behavior is not documented. Typically, identical image inputs at the same URL can be cached on the provider side, reducing token costs on repeated requests. Contact OrcaRouter support for specific details. Caching is model‑dependent and not guaranteed for preview models.
Token consumption for images is proportional to the number of 256×256 tiles needed to cover the image (after resizing). A 512×512 image uses 4 tiles (4 input tokens per tile? Not provided—exact formula depends on model). OrcaRouter may provide a token count in the API response usage field. Experiment with your own images to estimate cost per request.
Use the OpenAI‑compatible endpoint at https://api.orcarouter.ai/v1 with your API key. Set the model to "google/gemini-3-pro-image-preview". Format the request with a messages array containing both text and image parts. Images are passed as base64 data URLs or URLs with image_url objects. Example: {"model":"google/gemini-3-pro-image-preview","messages":[{"role":"user","content":[{"type":"text","text":"Describe this image"},{"type":"image_url","image_url":{"url":"data:image/png;base64,..."}}]}]}. Streaming is supported.
Standard OpenAI parameters: temperature (0–2), top_p, max_tokens (up to context window minus prompt tokens), stop sequences, frequency_penalty, presence_penalty. The model also accepts the "seed" parameter for deterministic outputs (not guaranteed). Parameter support is model‑dependent; some parameters may be ignored or have different default values. Test with your desired configuration.
Change your base URL from https://api.openai.com/v1 to https://api.orcarouter.ai/v1, update your API key to an OrcaRouter key, and change the model name to "google/gemini-3-pro-image-preview". Message structure (content array with text and image_url) is identical. If you use libraries like openai Python, just modify the base_url and api_key. Note: rate limits differ.
Authentication is via API key in the Authorization header (Bearer your_key). Rate limits are per‑key and depend on your plan. The API returns 429 when exceeded. There is no separate authentication for the model provider—OrcaRouter manages routing. For production, use a dedicated key and monitor usage in the OrcaRouter dashboard.
Both are multimodal (image+text in, text out). GPT‑4V is a mature production model with broader benchmark data. Nano Banana Pro is a preview; its true capabilities are less known. Context windows: GPT‑4V up to 128k vs 65k for this model. GPT‑4V supports higher resolution images. However, this model may offer lower costs and different reasoning strengths. Direct comparisons require task‑specific evaluation.
OrcaRouter offers multiple multimodal models (e.g., Claude 3 Vision, Llama 3.2 Vision). This Google preview provides a unique Gemini‑based architecture that may excel on certain Google‑centric tasks (e.g., understanding Google Docs screenshots). It has half the context window of some competitors. Pricing and latency vary; check OrcaRouter's comparison tables for per‑model rates.
The key advantage is native image input without needing a separate vision encoder. You can combine visual context with text in a single prompt. This reduces system complexity versus chaining two different models. However, text‑only models are cheaper and faster for tasks that don't need images. Choose based on whether the task requires visual understanding.
Gemini 2 Pro is a production model with long track record. This preview offers a glimpse of Gemini 3 Pro's architecture and may have different strengths (e.g., better handling of certain image types). However, it is a preview—stability and support are limited. For production deployments, Gemini 2 Pro (via OrcaRouter) is safer. Use this preview for early testing and feedback.
https://api.orcarouter.aimax_tokensresponse_formatseedstopstructured_outputstemperaturetop_p| Per request | $0.2400 |
| Currency | USD |
| Flat fee per API call (image generation models) | |
GET /api/public/models/google/gemini-3-pro-image-previewOpen @misc{orcarouter_gemini_3_pro_image_preview,
title = {Nano Banana Pro (Gemini 3 Pro Image Preview) API},
author = {Google},
year = {2025},
howpublished = {OrcaRouter},
url = {https://www.orcarouter.ai/models/google/gemini-3-pro-image-preview}
}Google. (2025). Nano Banana Pro (Gemini 3 Pro Image Preview) API. OrcaRouter. https://www.orcarouter.ai/models/google/gemini-3-pro-image-preview