Google launches Nano Banana 2.1: Image generation costs cut in half, benchmarks beat predecessor across the board

動區 BlockTempo · 2026-10-07T02:20:00.000Z

Google has released Nano Banana 2.1, an image generation model that outperforms both its predecessor and the higher-tier Pro version across the board in its own benchmarks. The API price per image has been nearly halved, while input and text output prices have gone up. The previous model will be shut down on October 29, leaving developers just 23 days to migrate. This article includes four sets of official examples. (Previously: Nano Banana 2 is here! Supports 4K and improved character consistency, making Gemini image generation faster and more precise) (More background: Google launches three new models, including Gemini 3.6 Flash! 3.5 Flash-Lite hits blazing speeds, and Gemini 4 pretraining has begun) The API price for a 1K image has dropped from $0.067 to $0.0336, nearly half. On the same day, Google also set a shutdown date for the previous model: October 29, just 23 days after the new version's launch. On October 6, Google announced Nano Banana 2.1, its new image generation model, in a post on its official X account. The company says it outperforms its predecessor across the board. Meet Nano Banana 2.1, our latest image generation and editing model This upgraded version outperforms our previous models across the board, with notable leaps in visual design, mask-based editing, and subject consistency to help you create more natural-looking images. pic.twitter.com/d858DrzrF3 — Google (@Google) October 6, 2026 It wins every benchmark—but Google is the judge According to the 10 benchmarks listed in the model card, with thinking enabled, 2.1 leads both its predecessor, Nano Banana 2, and the higher-tier Nano Banana Pro in every category. The thinking level determines how much internal drafting the model does before generating an image, with three settings: minimal, medium (the default), and high. Overall preference: 1050 points, compared with 990 for its predecessor and 935 for Pro. These scores are Elo ratings: people compare the work of two models side by side, pick the image they prefer, and the result is converted into points. A higher score means the model is chosen more often. Multi-subject consistency (whether the same people remain recognizable across different scenes) scored 1106, while the predecessor and Pro scored 978 and 1011, respectively. Mask-based editing (drawing around an area and asking the model to edit only that part) scored 1049, compared with 965 and 927 for its competitors. Infographic factuality is scored image by image by an automated grader, rather than through human evaluations, and the model card does not explain the score scale. 2.1 scored 0.521, nearly three times the predecessor's 0.179, and well ahead of Pro's 0.265 as well. With thinking turned off, 2.1 still scored 1015 in overall preference—higher than the predecessor's 990 with thinking enabled—so the improvement cannot be attributed entirely to “thinking one step further.” However, these are all Google-run benchmarks, the tables include only Google's own models, and there has been no third-party validation so far. In terms of specifications, the model supports up to 14 reference images, including up to 4 characters for character consistency and up to 10 objects for object recreation. It also fixes stitching artifacts in extremely tall or wide aspect ratios such as 1:4 and 8:1 at 2K and 4K resolutions. You don't need to use the API to try it: in addition to the Gemini API, the model card lists Gemini App, AI Mode in Google Search, Google AI Studio, Google Ads, Flow, and Stitch as launch platforms. Four sets of official examples Google has published sets of example images and prompts (instructions given to the model) on the official DeepMind showcase page. The first set focuses on visual design and typography: on the left is a “DESERT” off-road running poster covered in small text such as HEAT TOLERANCE, ZERO HORIZON, GPS coordinates, ELEV. 1,850 M, and a barcode; on the right is a 1970s travel poster. The prompt shown below the image asks for the giant “HORIZON” lettering to be partially obscured by a woman sitting on the roof of a vintage camper van. The notable point is that every specified string of text is spelled correctly, and the figure overlaps the lettering with a convincing sense of depth. ▲ The “DESERT” off-road running poster on the left and the “HORIZON” travel poster on the right; the prompt below corresponds to the image on the right. (Image source: Google DeepMind) Official prompt: Cinematic 1970s travel poster. A thin warm cream border frames the entire composition. Along the top header border is tiny, spaced, uppercase black text reading: “PACIFIC COAST ISSUE 08 NORTHERN SWELLS HIGHWAY ONE”. The background is a solid flat field of vibrant cobalt blue. Across the upper half, giant ultra-condensed cream-colored sans-serif text spells “HORIZON”. In the foreground, a young woman in a denim vest and sunglasses sits atop a vintage seafoam-teal camper van with surfboards mounted on top, overlapping and partially masking the giant letters to create depth. Rich film grain texture, warm natural lighting, high-contrast, ultra-sharp legible typography, no placeholder text, all te..."Google launches Nano Banana 2.1: Image generation costs cut in half, benchmarks beat predecessor across the board" This article was first published on BlockTempo (the most influential blockchain news media outlet in the Chinese-speaking world).

Google 推出生圖模型 Nano Banana 2.1，在自家評測中全面領先前代與定位更高的 Pro 版，API 單張圖價格約砍半，輸入與文字輸出卻漲價。前代型號 10 月 29 日關閉，開發者只有 23 天搬家，本文附四組官方範例。 （前情提要：Nano Banana 2 來了！支援 4K 強化角色一致性，Gemini 生圖更快更精準） （背景補充：Google 推出 Gemini 3.6 Flash 等三款新模型！3.5 Flash-Lite 飆出極速，Gemini 4 預訓練已啟動）   每張 1K 圖的 API 價格，從 0.067 美元降到 0.0336 美元，近乎腰斬。同一天，Google 也替前代型號訂下了關門日：10 月 29 日，距離新版上線後只留 23 天。 Google 10 月 6 日在官方 X 帳號發文推出生圖模型 Nano Banana 2.1，官方的說法是全面勝過前代。 Meet Nano Banana 2.1, our latest image generation and editing model This upgraded version outperforms our previous models across the board, with notable leaps in visual design, mask-based editing, and subject consistency to help you create more natural-looking images. pic.twitter.com/d858DrzrF3 — Google (@Google) October 6, 2026 跑分全贏，裁判是自己 根據模型卡列出的 10 項評測，開啟思考時，2.1 全數領先前代 Nano Banana 2，也領先定位更高的 Nano Banana Pro。思考等級是模型出圖前先在內部打草稿的力道，分 minimal、medium（預設）、high 三檔。 整體偏好 1050 分，前代 990，Pro 935。這裡的分數是 Elo：請真人把兩個模型的作品並排比較，選出較喜歡的一張，再換算成積分，愈高代表愈常被選中。 多角色一致性（同一批人物換了畫面，長相仍要認得出來）是 1106，前代與 Pro 分別是 978 與 1011。 遮罩編輯（用線條圈出範圍，要求模型只處理那一塊）是 1049，對手是 965 與 927。 資訊圖表事實性改由自動評分器逐張打分，不是真人評比，模型卡也沒有交代分數的刻度。2.1 拿到 0.521，前代是 0.179，接近三倍，連 Pro 的 0.265 也被拉開。 關掉思考，2.1 的整體偏好仍有 1015，高於開著思考的前代 990，進步不能全算在「多想一步」頭上。不過這些都是 Google 自己辦的評測，表格裡只有自家模型，目前也沒有第三方驗證。 規格上，參考圖最多 14 張，其中角色一致性最多 4 個角色、物件還原最多 10 個物件，並修掉 1:4、8:1 這類超長、超寬比例在 2K 與 4K 下的拼接瑕疵。 想試的人不必碰 API：模型卡列出的上線管道除了 Gemini API，還有 Gemini App、Google 搜尋的 AI Mode、Google AI Studio、Google Ads、Flow 與 Stitch。 官方展示的四組圖 Google 在 DeepMind 官方展示頁放了成組的示範圖與提示詞（prompt，也就是下給模型的指令）。 第一組是視覺設計與文字排版：左邊是越野跑主題的「DESERT」海報，印滿 HEAT TOLERANCE、ZERO HORIZON、GPS 座標、ELEV. 1,850 M 與條碼等小字；右邊是 1970 年代旅遊海報，附在圖下的是它的提示詞，要求巨大的「HORIZON」字樣被坐在復古露營車頂的女子部分遮住。看點是指定的每一串字都拼對了，人物壓在字上也做出了景深。 ▲ 左為「DESERT」越野跑海報，右為「HORIZON」旅遊海報，下方 prompt 對應右圖。（圖源：Google DeepMind） 官方 Prompt：Cinematic 1970s travel poster. A thin warm cream border frames the entire composition. Along the top header border is tiny, spaced, uppercase black text reading: “PACIFIC COAST ISSUE 08 NORTHERN SWELLS HIGHWAY ONE”. The background is a solid flat field of vibrant cobalt blue. Across the upper half, giant ultra-condensed cream-colored sans-serif text spells “HORIZON”. In the foreground, a young woman in a denim vest and sunglasses sits atop a vintage seafoam-teal camper van with surfboards mounted on top, overlapping and partially masking the giant letters to create depth. Rich film grain texture, warm natural lighting, high-contrast, ultra-sharp legible typography, no placeholder text, all te...〈Google 上線 Nano Banana 2.1：生圖成本砍半價、跑分全面勝過前代〉這篇文章最早發佈於動區BlockTempo《動區動趨-最具影響力的區塊鏈新聞媒體》。