Nano Banana 2.1 Arrives | API Pricing Cut in Half Compared to Previous Model
機械翻訳 / Machine-translated
A
AI Friends
5日前・16分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
If you use image-generation AI for work, have you ever thought, "I wish the cost per image were just a little cheaper"? Google's new model, "Nano Banana 2.1," is an update that directly addresses that concern. This article breaks down what has changed, what the new pricing looks like, and what to watch out for when switching over.
Nano Banana 2.1 is Google's image-generation AI capable of creating images from text and editing existing images. According to ITmedia NEWS, it was released on October 6, 2026 local time.
Google describes it as surpassing Nano Banana 2 in every respect. The areas of greatest improvement are visual design, mask-based editing (the ability to target and correct only a specific part of an image), and subject consistency (the ability to maintain the same appearance of a person or object across different images).
It will be rolled out across a wide range of platforms, including the Gemini app, Google Search's AI mode, Google AI Studio, Flow, Stitch, Google Ads, and the Gemini Enterprise Platform.
According to Android Authority, you can now pass up to 14 reference images in a single prompt. Output resolution goes up to 4K.
A bug that caused patterns to repeat like seams in landscape-oriented images (tiling artifacts) has apparently been fixed. Text rendering and infographic accuracy have also improved.
The model can maintain consistency for up to 4 people and up to 10 objects. This is useful when creating multiple images featuring the same character, or when swapping only the background in product photography.
Another noted improvement is that the original appearance is less likely to break down even after repeated rounds of editing.
In Google's internal evaluation, the overall image generation score rose from 990 in Nano Banana 2 to 1050. Decrypt reports that a factual accuracy metric also improved from 0.179 to 0.521.
That said, these are Google's own measurements. Third-party evaluations will be touched on in the competitive comparison section below.
For developers, pricing is the biggest point of interest. The output price per image is as follows:
A 1K image now costs about $0.0336 each. Generating 1,000 images would run roughly $33.60. Batch processing (submitting requests in bulk) is reported to reduce costs by an additional 50%.
For example, imagine someone running an online shop who generates 3,000 product images every month. At 1K resolution, that would cost about $100 per month — down from roughly $200 before.
One important caveat: input pricing has gone up. The cost per 1 million tokens (the unit AI uses to read text) has risen from $0.50 to $1.50.
Text and reasoning output is $7.50, and image output is $30. If your workflow involves passing many reference images along with lengthy prompts, the savings may fall short of that advertised "half price."
How does it stack up in third-party evaluations? According to CellCog's breakdown, on the image AI comparison site Arena, Nano Banana 2.1 ranked 5th (1,328 points) in the text-to-image category and 6th (1,428 points) in the image editing category. Both rankings are provisional.
GPT Image 2.5 Sunburst topped both categories with scores of 1,425 and 1,524, respectively. In terms of pricing, GPT Image 2.5 is listed at $0.013 to $0.053 per image.
In short, Nano Banana 2.1 represents a significant price drop within Google's own lineup — but it's hard to say it has matched the top competitors in either quality ranking or price. Choosing the right tool will depend on your specific use case.
The ITmedia NEWS article does not specify a Japanese domestic release date. However, this rollout includes the Gemini app and the developer-facing AI Studio. It is reasonable to expect that Japanese developers will also be able to try it via the API.
For Japanese businesses, the integration with Google Ads is significant. It has the potential to reduce the effort involved in creating banner ad assets.
The improvements to text rendering could also prove useful in situations where Japanese characters need to be embedded cleanly within images. However, accuracy with Japanese text has not been confirmed based on the information currently available. It is advisable to run small-scale tests before using it in actual business workflows.
For individual creators, the improvement in character consistency is a welcome addition. Imagine a doujinshi artist drawing each panel of a four-panel comic featuring the same protagonist — if the character's face or outfit changes from panel to panel, readers will be confused. Better consistency means less time spent on corrections.
According to CellCog's breakdown, Nano Banana 2 will be discontinued on October 29, 2026, and migration to the new model ID gemini-nano-banana-2.1 will be required. This end date comes from a single source in the information reviewed here. Be sure to verify it against official Google announcements.
If that date holds, there are roughly three weeks remaining. Anyone who has integrated the old model into their applications should verify functionality in a test environment as soon as possible.
Be aware of visual changes as well. Even with the same prompt, the style and feel of the output may differ. Companies with strict brand image guidelines are advised to compare outputs side by side before making the switch.
A. It was released on October 6, 2026 local time. It is being rolled out gradually across platforms including the Gemini app.
A. The image output pricing is approximately half the previous rate. The 1K image price dropped from $0.067 to $0.0336. However, input pricing has risen from $0.50 to $1.50.
A. Up to 14. Subject consistency is supported for up to 4 people and up to 10 objects.
A. Whether a free tier exists was not explicitly stated in the articles reviewed. Please check the official Gemini app terms for details.
A. It has been reported to end on October 29. However, sources are limited, so it is recommended to verify via official announcements.
Start by calculating your input and output costs based on your own usage patterns, and get your testing done before the old model's end date.
This article is a cross-post from AI Friends.
