Ideogram 4.5 對決 Nano Banana 2：實測 Ideogram 在自家發表影片中提出的說法

Ideogram 4.5 launched on 30 September 2026 with an unusual piece of marketing: it names the competition. The model page runs a side-by-side of the same edits applied to Ideogram 4.5, GPT Image 2.5 Sunburst, Nano Banana Pro and Nano Banana 2, the last of which is Nano Banana 2, the Gem​in 3.1 Flash Image checkpoint, in production since 26 February 2026. The verdict Ideogram prints under the video is blunt — "GPT Image and Nano Banana outputs become unusable within a few edits, while Ideogram 4.5 stays clean edit after edit." Naming a rival in your own launch asset is a commitment. It means the claim is checkable, and it means the model that gets checked is the one people already have wired into Gem​in, Search and Flow. This is what that claim looks like against the public record.

該主張，以及為支持它而提供的內容

撇除影片不談，這個主張有兩個部分，而它們的證據地位截然不同。

可量化的部分是漂移。Ideogram 的說法是，每一次編輯都會引入「像素位移、色彩變化與紋理偽影」，而 4.5「減少了這種漂移，在多輪編輯中保留細節」。這是一項基準測試可以評分的特性——依序套用同樣的六項編輯，測量每一輪像素偏離原始影像多遠——而目前沒有人這麼做。Artificial Analysis 在其編輯競技場中為單次編輯品質評分，並在其文字生圖競技場中為單次生成品質評分。多輪漂移不是這兩個排行榜上的任一軸線，這就是為什麼「幾次編輯後就不可用」這個說法，今日無法由這兩家實驗室以外的任何人查核。

可檢驗的部分在於，Ideogram 的比較是提示詞配對且不靠評審的。同樣的編輯套用到四個模型上，展示輸出結果，並要求讀者自行認同。這是一種合理的示範形式，也是多數圖像模型發表時所採用的論證方式。這並不是一種量測，而誠實的解讀是：Ideogram 展示了它的模型在自己挑選的提示詞集上，表現比三個具名競爭對手更經得起考驗——這有其價值，但與基準測試結果並不是同一回事。

Nano Banana 2 現在究竟是什麼

G​oogle's model is the incumbent in this matchup and it is worth stating precisely, because the Ideogram page compares against it while it has been in production for seven months.

Nano Banana 2 is the consumer name for Gemini 3.1 Flash Image. It shipped to the public on 26 February 2026 and reached general availability for developers with the model identifier gemini-3.1-flash-image; it is the default image generator across the G​emini app, G​oogle Search's Lens and AI Mode surfaces, Flow, and G​oogle Ads. That distribution is the part Ideogram's video does not engage with. Whatever the edit quality difference is, Nano Banana 2 is not a model you adopt — it is a model your users already have, and the question for a team is usually whether to pay for something better rather than whether to switch.

Its capability sheet is a generalist's: up to five consistent characters and fourteen objects in one prompt, 4K output with the extreme 1:4 and 1:8 ratio presets that banner formats need, in-image text translation into other languages with layout preserved, web-grounded generation, configurable thinking levels, SynthID watermarking and C2PA provenance metadata on every output. G​oogle prices it in output tokens, which resolves to roughly $0.067 per 1024×1024 image, about $0.101 at 2048×2048 and around $0.151 at 4K, with batch mode at approximately half and text input billed separately at $0.25 per million tokens.

兩個模型，逐行

• 已發布 — Ideogram 4.5 於 2026 年 9 月 30 日；Nano Banana 2（Gemini 3.1 Flash Image）於 2026 年 2 月 26 日

• 每張圖片價格 — Ideogram 4.5 從 $0.008（Very Low）到 $0.22（High，附來源圖片），中間分為 $0.03（Low）、$0.06（Medium）、$0.10（High）；Nano Banana 2 於 1024 約 $0.067、於 2048 約 $0.101、於 4K 約 $0.151

• Pricing basis — Ideogram charges per image by rendering speed and output size, with resolution from 1K to 2K changing nothing; G​oogle charges in output tokens, so cost scales with pixels delivered

• 獨立編輯排名 — Nano Banana 2 在 15,644 個樣本中以 Elo 1,108 排名第八；Ideogram 4.5 (High) 在 2,382 個樣本中以 1,064 排名第二十三

• 獨立文字轉圖像排名 — Nano Banana 2 以 Elo 1,125 位居第六，樣本數 18,454；Ideogram 4.5（High）以 Elo 1,014 位居第三十四，樣本數 2,247

• Demonstrated edit resolution — Ideogram shows a colourway edit inside a 4,016 × 6,016 pixel, 24.2-megapixel source with crop edges preserved for stitching back; G​oogle's ceiling is 4K output, which is a generation spec rather than an in-place edit claim

• Distribution — Nano Banana 2 is the default image engine of G​emini, G​oogle Search and Flow; Ideogram 4.5 is delivered through Ideogram's own app, its API endpoints and launch partners

• 參考圖片 — Nano Banana 2 最多可支援五個一致的角色與十四個物件；Ideogram 4.5 在所有方案層級皆接受來源圖片，High 以下不另收費

價格比較，妥善完成

粗略比較這兩份價目表會偏向 Ideogram 4.5，但認真比較則不會。

At the bottom of Ideogram's ladder a Very Low render costs $0.008 and a Low render $0.03 — both well under G​oogle's $0.067 at 1024. At the top, a High generation with a source image costs $0.22, or 3.3× G​oogle's 1024 price. So the answer to "which is cheaper" is entirely a function of which tier your pipeline can accept, and Ideogram's tiers are rendering speed, not capability levels. A team that can ship a Low render pays less than half of G​oogle's price for the same nominal 2K output. A team doing precision work on a real asset pays three times as much, and the resolution and drift claims are what they are buying with it.

The subtlety is what each vendor's price does as the image gets bigger. Ideogram charges the same $0.10 for High at 1K and at 2K, and the same $0.008 for Very Low at either size — output size is free within the tier. G​oogle charges in output tokens, so 4K costs roughly 2.25× the 1024 price. If your workload is large-format and you can live with a faster render, that inversion is where Ideogram 4.5's ladder earns its keep.

董事會，以及那個不討好任何人的數字

On Artificial Analysis's text-to-image leaderboard, where both models are measured on the same votes, Nano Banana 2 is sixth at 1,125 from 18,454 samples and Ideogram 4.5 (High) is thirty-fourth at 1,014 from 2,247. On the editing board the gap narrows: 1,108 against 1,064. G​oogle's error bars are tight and Ideogram's are not — 1,002 to 1,025 at the text-to-image end — which is what one day of voting looks like.

關於那點，有兩件事值得直說。首先，這個排序是真實的：在任一排行榜上，這些區間都不重疊。其次，它在一個要緊的方向上是暫時性的，因為一個上線僅一天的項目，其模型 Elo 通常會隨著票數累積而變動，而最初的樣本會偏向那些在發布當天有在關注的人。Ideogram 4.5 或許能縮小部分差距。但光靠它自己頁面上所做的那種比較，並不足以彌補二十三點的編輯 Elo 差距，不過這個數字尚未定論。

不討好任何人的數字，就是看板上的價格本身。Artificial Analysis 在編輯看板上列出 Ideogram 4.5（High）每千張圖片 $220——這正好是 Ideogram 自家 $0.22 的 High-with-source 方案——而 Nano Banana 2 為 $67。因此，這個獨立看板既認同兩份價目表，卻仍把較便宜的模型排在編輯軸的更高位置。這正是這場對決的關鍵：截至今日，經實測的編輯器是預算型的那款，而自稱編輯王者的模型，要付出三倍價格，換來被測得低它二十三分的結果。

打電話給他們其中一個

This is the rare case in this series where one of the two models is routable, and it is the one the subject model is measured against. G​oogle's image model is on OrcaRouter as google/gemini-3.1-flash-image-preview at $0.151 a call, provider list price passed through at 0% markup, so a G​oogle price change is live on our side the same day it happens. The routed identifier is G​oogle's preview snapshot; G​oogle's GA id for the same model is gemini-3.1-flash-image, and the rest of the G​emini image line is here too — google/gemini-3-pro-image-preview at $0.24, which is Nano Banana Pro — alongside the GPT-Image models and grok/grok-imagine-image, all behind one OpenAI-compatible endpoint.

Ideogram 4.5 並未收錄於目錄中。它可透過 Ideogram 自家的應用程式與 API 取得——端點為 /v2/image/generate/ideogram-4-5 以及 /v2/image/precise-edit/ideogram-4-5——也可透過首發合作夥伴取得；我們寧願這樣說，也不願暗示一條實際上並不存在的途徑。

這種安排之所以有用而非彆扭，在於這組對戰是你該親自跑的測試，而且它便宜的那一邊已經綁在同一把金鑰上。自動容錯移轉意味著，即使單一供應商在執行途中出狀況，指向 Nano Banana 2 的比較測試框架也不會整個垮掉；再加上路由 DSL 會把模型綁定在每個請求上，因此在你今天就能呼叫的模型、與另一個必須另行註冊才能用的模型之間做評比時，至少能讓基準維持不變。有項值得帶進那場測試的但書：編輯排行榜只評單次編輯，所以要正面對決、衡量 Ideogram 真正宣稱的事——圖像能否撐過連續六次編輯——這樣的測試框架得由你自己寫，而不是借用現成的。

哪一個該納入你的流程？

For most teams the answer is Nano Banana 2, and not because it is better at editing — because it is already everywhere. It is the default engine of the G​emini app and G​oogle Search, it takes five characters and fourteen objects in one prompt, it grounds on the web, it stamps SynthID and C2PA on output, and it is measured ahead of Ideogram 4.5 on both independent boards at three times less per image at the top of Ideogram's ladder. If your requirement is "generate this well, at volume, with provenance," the seven-month-old incumbent is the boring correct answer.

Ideogram 4.5 earns its place on one specific job: editing a real, large, existing asset in place. A 24-megapixel source edited at the crop boundary and stitched back without downsampling is not something Nano Banana 2 claims to do, and it is the operation Ideogram's $0.22 tier exists to price. If that is your workload, the premium is a real product decision rather than a markup. If it is not, you are paying three times G​oogle's price for a model currently measured lower.

能改變這個結論的，會是一項多輪漂移基準測試。Ideogram 在自家發布影片中点名 Nano Banana 2，並斷言其輸出在經過幾次編輯後就會崩解；這項斷言是價格溢價的全部依據，它是可驗證的，而沒有任何獨立機構對其進行測試。在有機構這麼做之前，誠實的立場是：這項說法聽來合理、尚未獲得證實，而且針對的正是你如今已可從我們型錄中呼叫的那個確切模型。

The rest of the G​emini image line is here too, behind one OpenAI-compatible endpoint — google/gemini-3-pro-image-preview at $0.24, the GPT-Image models and grok/grok-imagine-image, on the same key.