Genie 4 與 Google Veo 3.1 的標題卡,標題為「一個你走進去的世界,還是一段你剪出來的片段」,其中一張卡將 Genie 4 描述為 720p、約一分鐘的互動世界,另一張則將 Google Veo 3.1 描述為最高 4K、長 4 到 8 秒的渲染片段。
Guides & Insights

Genie 4 對決 Google Veo 3.1:是讓你漫步其中的場所,還是剪輯出來的短片?

作者

Gideon Frost

發佈日期

最新模型 · 20查看全部模型 →
基準測試:Artificial Analysis · 每日更新
返回全部文章

Sixty seconds of output costs $24 on G​oogle Veo 3.1 at its standard tier, $6.00 on Veo 3.1 Fast, and $3.00 on Veo 3.1 Lite. Sixty seconds of Genie 4 costs $99.99 a month, because there is no sixty seconds to buy — the model card caps a session at around a minute of continuous interaction, and the only way in is a G​oogle AI Ultra subscription. Put those two prices next to each other and the comparison looks absurd, which is the point: Genie 4 and G​oogle Veo 3.1 are not two versions of the same purchase. Veo renders a shot you will cut. Genie generates a place you will walk through while it is still being generated. Every other difference between them follows from that one.

Both went through a status change this autumn, and the two changes point in opposite directions. Veo 3.1 lost its consumer-facing developer home: the Veo 3.1 Preview in the G​emini API was shut down on 22 October 2026, and G​oogle now points developers at G​emini Omni 1.2 Flash or the GA Veo models on the G​emini Enterprise Agent Platform. Genie 4 gained one: announced on 27 October 2026 to trusted testers, it opened to eligible G​oogle AI Ultra subscribers inside Project Genie on 18 November, with a model card that finally published a resolution and a frame rate. One model got narrower and more expensive to reach; the other got wider and, for the first time, measurable.

先談分類法,因為它決定之後的一切。

影片模型是一個函式。你給它一個提示詞,可選地加上第一幀或參考圖像,四到八秒後{{1}}你就會得到一個檔案{{/1}}。你可以從佇列並行呼叫它一千次。輸出在它存在的那一刻就已完成。

世界模型是一個過程。它會回應你的動作,逐幀渲染,只要工作階段持續多久就持續多久。除非你錄製下來,否則最後不會有任何檔案。輸出永遠不會完成——它正在你所在位置前方幾幀之處被創造出來。

That distinction determines what each is even capable of. V​eo 3.1 can produce a 4K shot of a rain-soaked street that a colourist will accept. It cannot let a director say "make it night" and keep walking down the same street. Genie 4 can do exactly that — mid-session prompts change weather, add characters, redirect action — at 720p and up to 24 frames a second, for about a minute, and then the session ends.

實際存在的數字

G​oogle published far more about Veo 3.1 than about Genie 4, and the asymmetry is itself a finding. Every Veo figure below is from G​oogle's own documentation; every Genie figure is from the model card published on 18 November 2026. Neither vendor has published a benchmark that a third party has reproduced.

• 輸出 — Genie 4 為 720p、最高 24 fps、即時生成,並附同步音訊;相較之下,Veo 3.1 為 720p、1080p 或 4K,同樣是 24 fps,但採離線渲染。

• 長度 — Genie 4 每次工作階段約可連續互動一分鐘,相較之下,Veo 3.1 每個片段為 4、6 或 8 秒,並可透過後續多次處理延伸。

• 輸入 — Genie 4 接受文字、一張圖片,以及即時鍵盤與滑鼠導覽,並可在工作階段中途接受新的文字提示;相對地,Veo 3.1 接受文字、第一或最後一個影格、最多三張用於主體與風格的參考圖片,以及鏡頭與動作控制。

• 音訊 — 兩者皆可原生生成。Genie 4 會產生環境音、音樂,以及隨著場景變化調整的互動式對話;Veo 3.1 則會產生與影片片段相連的音效、環境音與對話,並提供預設為關閉的音訊開關。

• Price — Genie 4 is $99.99 a month for G​oogle AI Ultra, with a $199.99 tier that raises usage limits rather than capability vs Veo 3.1 at $0.40 per second of video with audio at 720p or 1080p, $0.60 at 4K, $0.20 and $0.40 without audio; Fast at $0.10, $0.12 and $0.30 with audio; Lite at $0.05 and $0.08 with audio.

• Access — Genie 4 is a browser product for Ultra subscribers, with an AI Pro waitlist, and has no API and no model ID vs Veo 3.1 is a billed API on the G​emini Enterprise Agent Platform, no longer reachable through the G​emini API's preview endpoints.

• 來源溯源 — 兩者皆以 SynthID 加上浮水印。

• Independent score — neither has one for this comparison. G​oogle's Veo 3.1 preference study is Google-run, on 1,003 MovieGenBench prompts, and predates this release cycle. Genie 4 has no published evaluation of any kind.

每一項每分鐘實際上要花多少

這正是這兩款產品不再能相提並論的地方,值得實際算一算,而不是憑感覺。

A one-minute Veo 3.1 sequence at standard quality with audio is $24.00, on G​oogle's published Vertex AI rate card of $0.40 per second at 720p or 1080p — $0.60 at 4K, and $0.20 and $0.40 with the audio track switched off. At Fast it is $6.00 a minute; at Lite $3.00. But Veo does not make one-minute clips — the longest single generation is eight seconds — so that minute is really eight generations plus whatever scene extension costs, and continuity across the joins is something you manage in the edit rather than something the model maintains. For a twenty-second insert in a brand film, that discipline is exactly what you want.

Genie 4 at $99.99 a month buys sessions, not seconds. Keep one open all afternoon and the marginal cost of the tenth session is zero; the constraint is the usage quota of your Ultra tier, which G​oogle has not published in units anyone can plan against. If your work is exploratory — blocking a scene, testing whether a set reads at eye level, showing a client what a space feels like before anyone builds it — the subscription model is dramatically cheaper than renting seconds. If your work is delivery, it is not a model of the thing you need at all.

尚未解決的部分是解析度。Veo 3.1 在最高階方案中提供原生 4K。Genie 4 則提供 720p。720p 的即時擷取屬於預視覺化素材,而非母帶;將它升頻並放進 4K 時間軸,是在明知這些像素被憑空生成過兩次的情況下所做的決定。

當推理層是唯一可共享的部分時

G​oogle's Genie 4 architecture is a stack, not a single model: G​emini Omni Flash grounds the visuals, and a G​emini model runs in the loop to plan steps and check frames against goals. G​oogle names Gemini 3.8 Flash in that role.

That layer is the one piece of this comparison you can actually call today on a general-purpose API. Gemini 3.8 Flash is a live endpoint on our catalogue as google/gemini-3.8-flash, at $0.75 per million input tokens and $3.75 per million output, provider list price passed through with no markup on top — so a G​oogle rate change lands on your bill the same day it lands on G​oogle's. If you are building the loop around a world model rather than the world model itself — decomposing a goal, writing the steer prompts, verifying the frame you got against the frame you wanted — that is an integration you can finish this week, with automatic failover behind one key if a provider wobbles.

Screenshot of the OrcaRouter model page for google/gemini-3.8-flash rendered in English, showing the model identifier and its per-million-token input and output pricing.

For the record, and because the question comes up: neither Genie 4 nor Veo 3.1 is an endpoint on our platform. Genie 4 is not an endpoint anywhere. Veo 3.1 is a G​oogle Cloud model with a G​oogle Cloud contract. We are not going to pretend otherwise to make a routing story work.

誠實的裁決

當交付成果是檔案時,就選 Veo 3.1:一個鏡頭、一段插入鏡頭、一支社群剪輯,任何必須是 1080p 或 4K、經過調色,並交給剪輯師的內容。它有公開的價目表、有文件記載的 API、企業級配套,背後還有一年的製作實績。如果你當初是針對那些端點開發,10 月 22 日的預覽版關閉是實實在在的成本,但正式可用(GA)的路徑存在,而且有文件記載。

Choose Genie 4 when the deliverable is a decision: which of three layouts reads better, whether a set works from a child's eye level, what a space feels like before it is built. One minute at 720p is enough to answer those questions and far too little to answer a delivery question. The gallery — with remixable worlds and the two named collaborations, "Villa Giallo" and "The Life of Ants" — is the clearest statement of what G​oogle thinks the product is for, and none of it is a render farm.

Screenshot of Google AI Studio documentation stating that video generation in the Gemini API, covering the Veo 3.1 preview endpoints, was shut down.

能改變答案的,是提供按秒計價的 Genie 4 API,或是接受在片段中途變更指令的 Veo。兩者任一都會讓差距消失。在那之前,這項比較與其說是正面對決,不如說是兩款恰好共用同一家供應商、相同影格率與浮水印的工具。

Generated two-column scoreboard comparing Genie 4 and Google Veo 3.1 across output, length, input, audio, price and access, with Genie 4 at 720p and about a minute for $99.99 a month in a browser and Veo 3.1 at up to 4K for 4 to 8 seconds at $0.40 a second on the billed API.