Google has released the Nano Banana 2.1 model for high-resolution image generation and editing at half the cost of the previous version. The new model is based on the Gemini 3.6 Flash architecture, offering qualitative leaps in visual consistency and graphic mask processing.
- Cost reduction and Gemini Flash architectural upgrade
- Pricing details and newly introduced engineering features
- Human performance tests and competitive platform rankings
- Deployment roadmap, technical constraints, and shutdown dates
- Frequently asked questions
Cost reduction and Gemini Flash architectural upgrade
Google officially announced on October 6, 2026, the launch of its latest visual artificial intelligence model named “Nano Banana 2.1”, a model dedicated to generating and editing digital images based directly on the ultra-efficient “Gemini 3.6 Flash” architecture. The most prominent surprise of this launch is the imposition of operating fees equal to half the previous cost established in the second generation, giving developers and startups an exceptional opportunity to reduce visual computing expenses while benefiting from qualitative leaps in design quality, masking, and the appearance consistency of illustrated elements and characters.
This advanced release replaces the previous Nano Banana model launched last February, with Google setting a final deadline ending on October 29 to permanently shut down the old model, requiring developers to rush to update their APIs. The company confirmed via its official accounts that the new model outperforms its predecessors across all benchmark metrics, especially in integrating complex texts within designs and removing grid-repetition flaws that appeared in wide panoramic images.
Pricing details and newly introduced engineering features
Gemini API pricing documentation shows that the cost of image generation with the new Nano Banana model has settled at $30 per million output tokens. This figure practically translates to a cost of approximately 3.4 cents for 1K resolution images, 5 cents for 2K, and 7.6 cents for 4K ultra-high-resolution images, compared to previous prices of 6.7, 10.1, and 15.1 cents for the same dimensions. Despite the increase in input token costs to $1.50 per million tokens compared to $0.50 previously and the absence of a free tier, the batch processing option makes it possible to cut the total image cost in half once again.
On the functional front, Google added tangible improvements to infographic layout, allowing the use of up to 14 reference images in a single prompt to guide the visual style, along with equipping the model with three adjustable levels of deliberate “thinking patterns” before generation. Although Google offers the model as an economical sibling to the “Pro” model, independent field tests indicate that the professional version still produces superior realistic details in complex scenarios.
Human performance tests and competitive platform rankings
Google showcased the results of human preference evaluations in the model specification card, where Nano Banana in thinking mode recorded a benchmark rating of 1,050 points according to the Elo rating system, clearly outperforming the second generation which achieved 990 points and the professional version which garnered 935 points. However, independent evaluations across global platforms were less enthusiastic than the company’s internal estimates; the general leaderboard of the Genki AI platform placed the new model in fifth place globally on its launch day, trailing OpenAI’s “GPT Image 2.5” models and Microsoft’s “MAI” model.
This discrepancy is due to different evaluation criteria, as independent developers focus on strict accuracy in executing complex compound instructions and avoiding facial distortion at sharp angles, while corporate tests focus on the overall balance between processing speed, color aesthetics, and overall lighting in simple everyday scenes.
Deployment roadmap, technical constraints, and shutdown dates
Google has begun making the new model available across a wide ecosystem of its services, including the Gemini chat app, the search engine’s AI mode, Google AI Studio, Flow, Stitch, and Google Ads platforms, alongside the enterprise platform for businesses. Robbie Stein, Vice President of Product for Search, stated that the model has actually begun operating in smart search interfaces, noting that general generative summaries have not fully adopted the new release yet.
Technical documentation acknowledges some shortcomings that are still under improvement, such as visual element hallucination in crowded scenes, blurring of small-sized texts, minor differences in character feature consistency upon repeated generation, as well as positional errors in mask-based editing. Experts also spotted a technical conflict in the documentation regarding context window capacity, which was listed as one million tokens in the model card while API specifications set it at approximately 131,072 tokens, requiring developers to conduct thorough software tests before the final shutdown date on October 29.
Frequently asked questions
Question: How much does image generation cost with the new Nano Banana model?
Answer: The cost is $30 per million output tokens, which equals 3.4 cents for 1K resolution images and about 7.6 cents for 4K images.
Question: What is the main upgrade in the model’s software architecture?
Answer: The model was built on the Gemini 3.6 Flash architecture, with support for multiple thinking levels and the use of up to 14 reference images for guidance.
Question: When will older visual versions of the API be discontinued?
Answer: Google has set October 29, 2026, as the final deadline to discontinue the previous Flash Image model, urging developers to migrate to the updated version.
