Nano Banana 2 vs Pro: 30 paired images, timing and cost

Magic Hour publishes this benchmark and supplied the platform route and credits used for the paired runs. The retained 30 images, prompts, timings, and recorded costs support the findings below; the result is limited to the tested model versions and settings rather than a universal quality ranking.

What did we measure?

Five prompts covered a Lisbon street scene, typography, a five-subject composition, watercolor illustration, and a labeled diagram. Each prompt ran three times per model at 1K, 1:1, one image per request. Requests were serial and interleaved by model. Three additional Nano Banana 2 requests used 640px, and one request used the default router: 34 requests in total, all completed.

The collector polled every two seconds. Reported time includes the Magic Hour request path, queue, generation, and polling delay; it is not raw Gemini inference latency. The public benchmark repository contains the prompts, code, request log, responses, and output images.

Timing results at 1K

Pro was slower in 14 of the 15 paired requests. The single exception was a photoreal repeat: 27.3 seconds for Nano Banana 2 and 25.6 for Pro. These observations describe one day's small sample, not a service-level promise.

Every paired request

Does the prompt change the gap?

The prompt-level ratios ranged from 1.46 to 2.43 in these three-repeat groups. That is too little evidence to predict the speed gap for a future workload from its subject matter. The three 640px Nano Banana 2 runs averaged 17.3 seconds; this small sample does not establish a resolution-speed relationship.

Actual images from the first paired attempt

The images below are prompt 1, repeat 1, chosen because they are the first pair in the log—not because they were selected as the best outputs. The prompt requested a photorealistic Lisbon street scene with a yellow Tram 28, cobblestones, balconies, laundry, and golden-hour light.

Nano Banana 2: 15.2 seconds, 100 credits

Nano Banana Pro: 29.6 seconds, 150 credits

Inspect composition, requested details, text, and unwanted additions in the full image collection. The study has no completed human quality scorecard. A preference for this pair is not a general ranking, and the generation prompts do not evaluate localized image editing.

What did each image cost?

Every measured 1K Nano Banana 2 request charged 100 Magic Hour credits; every 1K Pro request charged 150. The three 640px Nano Banana 2 requests also charged 100 each. These are the recorded July 21 charges, not a flat rate for every output size, provider, or future request.

The historical comparison chart below should be read with that scope. Its Google price figures describe a different provider and are not measured Magic Hour invoices.

For current purchasing, check Magic Hour's model reference and the quote for your request. The reference checked September 9, 2026 lists Nano Banana 2 from 100 credits and Pro from 150; higher resolutions and additional images can change the charge. Nano Banana 2 supports 640px, 1K, 2K, and 4K; Pro starts at 1K. The current catalog lists both on paid Creator, Pro, and Business tiers. The benchmark's earlier account restriction does not establish today's plan-wide resolution limits.

For Google's standard direct-API billing, use the Nano Banana Pro pricing guide and its primary pricing source. Keep input charges, output size, billing mode, and provider distinct.

How should you choose for your workload?

Record accepted outputs and rejection reasons. A lower charge per candidate is useful only if enough candidates meet your brief. Do not infer that underlying Gemini features such as document input or search grounding are exposed by a third-party endpoint; use that endpoint's documented fields.

Use both models through one API

The measured requests used Magic Hour's image-generation API. The paired requests kept the prompt, image count, aspect ratio, and resolution fixed and changed the model field from nano-banana-2 to nano-banana-pro.

Use Magic Hour's image generator for an interactive comparison. Choose Pro when its observed output better satisfies your brief, not because this timing study proved an accuracy advantage. Verify the current quote before increasing resolution or batch size.

Existing readers using the article's API signup link should confirm any promotional offer at checkout; this benchmark does not verify a current discount.

Does Nano Banana 2 replace Nano Banana Pro?

They are separate choices: Nano Banana 2 corresponds to Gemini 3.1 Flash Image and Nano Banana Pro to Gemini 3 Pro Image. The current Magic Hour catalog lists both. This study supports a narrower conclusion: Nano Banana 2 was faster and charged fewer credits in the tested 1K requests. It does not establish that one replaces the other for every task.

Sources and verification

The request log, exact prompts, and all output images preserve the experiment. The timing summary was recomputed from the published log during this September 9, 2026 update. No new generations or human quality scores were added.