# Veo 3 > Source: https://aiwiki.ai/wiki/veo_3 > Updated: 2026-10-30 > Fact-checked: 2026-10-30 > Categories: AI Models, Google DeepMind, Video Generation > License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)" > Cite as: AI Wiki. "Veo 3." aiwiki.ai, 30 Oct 2026. https://aiwiki.ai/wiki/veo_3 > From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution. **Veo 3** is a video generation model developed by [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind) and announced at Google I/O on May 20, 2025. It generates video and a synchronized soundtrack together, including spoken dialogue, sound effects, ambient noise, and music, rather than producing silent footage that has to be scored afterwards.[3][28] Google described it as its "first video model to incorporate high-fidelity video outputs and native audio," and launch coverage singled out the audio as the main thing separating it from rivals such as [OpenAI](https://aiwiki.ai/wiki/openai)'s [Sora](https://aiwiki.ai/wiki/sora).[2][4] It is the third major iteration of the [Veo](https://aiwiki.ai/wiki/veo) family.[30] Through the Gemini API, Veo 3 produced clips of up to eight seconds at 720p or 1080p and 24 frames per second until its endpoints there were shut down in 2026.[22][33][40]
Google DeepMind logo
Google DeepMind logo. Source: Wikimedia Commons, Public domain.
The model has been offered through several [Google](https://aiwiki.ai/wiki/google) products: the Flow filmmaking tool, the Gemini app, [Vertex AI](https://aiwiki.ai/wiki/google_vertex_ai), [Google AI Studio](https://aiwiki.ai/wiki/google_ai_studio), and the Gemini API.[1][2][26] Consumer access at launch was limited to Google AI Ultra subscribers in the United States, at $249.99 per month.[4][5] Google followed it with [Veo 3.1](https://aiwiki.ai/wiki/veo_3_1) on October 15, 2025, which added reference images, first-and-last-frame control, and scene extension, all with audio.[13][53] On May 19, 2026, Google announced [Gemini Omni](https://aiwiki.ai/wiki/gemini_omni), a multimodal video generation and editing model that replaced Veo in the Gemini app, and the Veo 3.1 preview endpoints in the Gemini API were shut down on October 22, 2026.[8][38][39] Google DeepMind CEO [Demis Hassabis](https://aiwiki.ai/wiki/demis_hassabis) told a press briefing at the launch, "For the first time, we're emerging from the silent era of video generation," a line that was widely repeated in coverage.[41][46] ## Overview Veo 3 sits at the intersection of two research directions in generative media: [latent diffusion model](https://aiwiki.ai/wiki/latent_diffusion)s for video synthesis, and joint audio-visual generation. According to Google's technical report, the diffusion process is applied jointly to temporal audio latents and spatio-temporal video latents, so sound and picture come out of the same model rather than being matched up in [post-processing](https://aiwiki.ai/wiki/post-processing).[23] At launch, audio was the headline difference between Veo 3 and the leading commercial video generators. TIME wrote that "unlike the outputs of previous video generators like OpenAI's Sora, Veo 3 videos can include dialogue, soundtracks and sound effects."[9] Audio-generating tools themselves were not new; TechCrunch noted that "AI-powered sound-generating tools aren't novel," and Google's own claim was that Veo 3 could understand the raw pixels of its videos and sync generated sound to them automatically.[41] The model launched into a crowded field. CNBC described Veo 3 as a competitor to OpenAI's Sora video generator.[4] Runway had released [Runway Gen-4](https://aiwiki.ai/wiki/runway_gen_4) in March 2025, [Kuaishou](https://aiwiki.ai/wiki/kuaishou) had iterated through several Kling versions, [MiniMax](https://aiwiki.ai/wiki/minimax) was selling its [Hailuo](https://aiwiki.ai/wiki/hailuo) models, and [Alibaba](https://aiwiki.ai/wiki/alibaba) had released the open-weight [Wan 2.1](https://aiwiki.ai/wiki/wan_2_1) in February 2025.[66][69][71] In Google's own launch evaluation on Meta's MovieGenBench prompts, Veo 3 was compared against Meta's [Movie Gen](https://aiwiki.ai/wiki/movie_gen), Kling 2.0, MiniMax, Sora Turbo, Runway Gen-3, and Wan 2.1, and for the audio tasks against Kling 2.0 paired with the MMAudio sound model.[15] Google also built Veo 3 into its own products, including the Gemini app, Flow, and Vertex AI, and later YouTube Shorts.[1][5][48] Over the following year the field changed quickly. OpenAI released [Sora 2](https://aiwiki.ai/wiki/sora_2), which also generates synchronized audio, on September 30, 2025, and Kuaishou added what it called "simultaneous audio-visual generation" to Kling with Kling 2.6 in December 2025.[61][64] Google iterated on its own model with Veo 3.1 in October 2025 and a January 2026 update, and then moved its consumer video generation to Gemini Omni in May 2026.[12][13][38]
"Meet Veo 3, our latest video generation model", official video from Google, published May 30, 2025. Watch on YouTube.
## History ### Veo 1 The first version of the Veo model was announced at Google I/O in May 2024.[30] Google said it could generate 1080p video clips over a minute in length from text or image prompts.[30] Veo 1 was made available to a limited set of creators and filmmakers through a waitlist for the VideoFX tool in Google Labs, and an earlier generation of Veo was later used in the Dream Screen background generator in YouTube Shorts.[24][27] It did not generate audio. Veo 1 was announced three months after OpenAI first showed its original Sora model in February 2024.[61] ### Veo 2 In December 2024, Google released [Veo 2](https://aiwiki.ai/wiki/veo_2), available initially through the VideoFX experimental platform.[27] Google said Veo 2 had an improved understanding of real-world physics and of human movement and expression, understood "the unique language of cinematography," and could produce video "at resolutions up to 4K, and extended to minutes in length."[27] Veo 2 was later made available on Vertex AI and in the Gemini app. It still generated video only, without audio.[30] When Flow launched in May 2025, Google AI Pro subscribers used it with Veo 2, while Veo 3 was reserved for Ultra subscribers.[5][26] ### When was Veo 3 released? Veo 3 was announced on May 20, 2025, at Google I/O, where Google said the model "now has native audio generation."[28] Its central advance over Veo 2 was native audio-visual generation: Google's technical report describes a latent diffusion model in which "the diffusion process is applied jointly to the temporal audio latents, and the spatio-temporal video latents."[23] The model became available the same day in the Gemini app and in Flow for U.S. subscribers to Google AI Ultra, and in private preview on Vertex AI.[1][4][5] Veo 3 opened to all Google Cloud customers in public preview on Vertex AI on June 27, 2025, and developer access through the Gemini API followed on July 17, 2025, as a paid preview at $0.75 per second of video with audio.[2][57][60] The announcement drew immediate public attention, partly because of how quickly creators began sharing viral outputs, including man-on-the-street interviews, mock advertisements, imagined film trailers, and ASMR clips.[9][46] The Will Smith eating spaghetti test, an informal benchmark that began with a distorted AI-generated clip in March 2023, was re-run with Veo 3 within two days of the launch and widely shared as a measure of how far the technology had come.[18][19] Google expanded access in stages. On May 30, 2025, it launched the Ultra plan in the United Kingdom, said Ultra was available in 73 countries, and gave Google AI Pro subscribers in those countries limited Veo 3 access in Flow and a 10-video trial in the Gemini app, including on mobile.[16][29] In July 2025 Google said it had expanded Veo 3 access to Pro subscribers in more than 150 countries and that users had generated more than 40 million Veo 3 videos in the Gemini app and Flow over the previous seven weeks.[45] ### Veo 3 Fast Veo 3 Fast is a lighter-weight variant optimized for speed and cost. It reached the Gemini app and Flow on June 9, 2025, when Google described it as twice as fast to generate as Veo 3 while keeping 720p resolution; Google AI Pro subscribers received three Veo 3 Fast generations per day in the Gemini app, and a Veo 3 Fast generation in Flow cost 20 credits.[44] On July 31, 2025, Veo 3 Fast was added to the Gemini API at $0.40 per second of video with audio, against $0.75 for the standard model, together with image-to-video support for both models.[32][74] In an announcement dated September 8, 2025, and recorded in the Gemini API changelog on September 9, Google declared Veo 3 and Veo 3 Fast generally available in the Gemini API, added 1080p output and the 9:16 vertical aspect ratio to both, and cut prices: Veo 3 with audio fell from $0.75 to $0.40 per second, and Veo 3 Fast from $0.40 to $0.15 per second.[20][21][33] In July 2025 the Gemini API also listed a $0.50 per second price for Veo 3 video without audio, charged only when the audio track was blocked by safety filters; from August 2025 it listed only prices for video with audio.[34] Vertex AI listed separate, lower prices for video without audio, and by October 2025 Veo 3 cost $0.20 per second and Veo 3 Fast $0.10 per second without audio there.[7] ### Veo 3.1 Veo 3.1 was released on October 15, 2025, as an update to Veo 3.[13][53] Google said it "brings richer audio, more narrative control, and enhanced realism that captures true-to-life textures," with "stronger prompt adherence and improved audiovisual quality when turning images into videos."[13] It became available the same day in the Gemini app, Flow, the Gemini API (as a paid preview, alongside Veo 3.1 Fast), and Vertex AI.[13][53] In the Gemini API, Google said "Veo 3.1 is the same price as Veo 3."[53] The release brought audio, "for the first time," to three existing Flow capabilities: Ingredients to Video, which uses multiple reference images to control characters, objects, and style; Frames to Video, which generates a transition between a start and an end image; and Extend, which continues a clip into videos "lasting for a minute or more," generating each new segment from the final second of the previous one.[13] In the Gemini API, developers could supply up to three reference images.[53] By the time of the release, Google said more than 275 million videos had been generated in Flow.[13] ### Veo 3.1 January 2026 update On January 13, 2026, Google updated Veo 3.1's Ingredients to Video mode. The update added native 9:16 vertical generation to Ingredients to Video, improved identity consistency so characters keep the same appearance across scenes, and introduced "state-of-the-art upscaling to 1080p and 4K resolution."[12][35] Google's announcement described the 1080p and 4K options as upscaling, although the Gemini API documentation said Veo 3.1 "can also directly generate 720p, 1080p or 4k videos"; the options were available only through Flow, the Gemini API, and Vertex AI, not in the consumer Gemini app or YouTube.[12][35][75] The same update brought Ingredients to Video to YouTube Shorts and the YouTube Create app for the first time.[12] ### Veo 3.1 Lite On March 31, 2026, Google released Veo 3.1 Lite in the Gemini API and Google AI Studio, calling it its "most cost-effective video model" and saying it came "at less than 50% of the cost of Veo 3.1 Fast, but with the same speed."[33][37] It supports text-to-video and image-to-video in 16:9 and 9:16 at 720p or 1080p, with clips of 4, 6, or 8 seconds, and does not offer 4K output.[34][37] The Gemini API priced it at $0.05 per second at 720p and $0.08 per second at 1080p.[34] In the same announcement Google said it would cut the price of Veo 3.1 Fast on April 7, 2026; the Gemini API pricing page then listed Veo 3.1 Fast at $0.10 per second at 720p, $0.12 at 1080p, and $0.30 at 4K, down from $0.15 (and $0.35 at 4K).[34][37] Veo 3.1 Lite entered public preview on Vertex AI on April 2, 2026.[57] ### Successor: Gemini Omni At Google I/O on May 19, 2026, Google announced Gemini Omni. Koray Kavukcuoglu, chief technology officer of Google DeepMind, introduced Gemini Omni Flash as "a model that can create anything from any input," starting with video.[38] Omni accepts combinations of images, audio, video, and text as input and lets users edit generated videos through conversation.[38] Google rolled out Gemini Omni Flash to Google AI Plus, Pro, and Ultra subscribers in the Gemini app and Google Flow, and, starting that week, at no cost in YouTube Shorts and the YouTube Create app.[38] Google's Gemini app video page states that "Gemini Omni will replace Veo in the Gemini app."[39] Flow continued to offer Veo 3.1 Lite, Fast, and Quality models alongside Gemini Omni Flash.[6][56] For developers, Gemini Omni Flash entered public preview in the Gemini API on June 30, 2026, and became generally available as `gemini-omni-1.1-flash` on August 27, 2026.[33] Google then retired the Veo endpoints in the Gemini API. The Veo 3 preview models had already been shut down on November 12, 2025; the stable `veo-3.0-generate-001` and `veo-3.0-fast-generate-001` models were shut down on June 30, 2026; and all Veo 3.1 preview models, including Lite, were shut down on October 22, 2026, with `gemini-omni-1.1-flash` listed as the replacement.[8][40] On Google Cloud, where Vertex AI became part of the Gemini Enterprise Agent Platform on April 22, 2026, the generally available Veo 3 and Veo 3 Fast models were deprecated on March 24, 2026 and retired on June 30, 2026, with Veo 3.1 as the recommended replacement; the Veo 3.1 models that became generally available there on November 17, 2025 carried a retirement date of "November 17, 2026 or later."[57][58] As of October 2026, Google DeepMind's Veo page still presented Veo 3.1 as the current Veo model.[3] ## How does Veo 3 generate audio? The defining feature of Veo 3 is that it generates audio together with the video.[3] Google's model page says Veo 3 "lets you add sound effects, ambient noise, and even dialogue to your creations," generating all of the audio natively, and the Gemini API launch post described it as natively generating "dialogue, effects, and music" and synchronizing them with video "in a single pass."[2][3] Earlier commercial video generators, including Veo 2 and the original Sora, produced silent output, and audio had to be added in post-production.[9][14] Veo 3 generates three broad categories of audio: - **Dialogue**: Characters speak lines given in the prompt, with mouth movements synchronized to the generated speech.[2][3] - **Sound effects**: Actions in the scene trigger matching sounds, such as footsteps, impacts, rain, or engine noise.[2][3] - **Ambient and background audio**: The model generates environmental sound consistent with the setting, and music where the prompt calls for it.[2][3] Prompts can describe the audio directly. Google's own examples include lines of dialogue placed in quotation marks and explicit cues such as "Audio: wings flapping, birdsong," alongside the visual description.[3] When a prompt does not specify sound, the model fills in audio consistent with the scene. The quality of the audio varies. Early users noticed that the Will Smith spaghetti clip had loud crunching and squelching sounds even though spaghetti does not crunch, which Ars Technica attributed to training data containing many clips of chewing paired with exaggerated sound effects.[18] Africa Check found that Veo mispronounced words in tonal African languages, such as the Yoruba words "amala" and "babalawo."[14] Through the Gemini API, audio is always on for Veo 3 and Veo 3 Fast.[22] ### Technical architecture Google has published a short technical report and a [model card](https://aiwiki.ai/wiki/model_card) for Veo 3 rather than a full research paper.[15][23] According to the technical report, Veo 3 is a latent [diffusion model](https://aiwiki.ai/wiki/diffusion_model). Video and audio are encoded by separate autoencoders into compressed latent representations, "in which learning can take place more efficiently than with the raw pixels or waveform."[23] During training, a transformer-based denoising network learns to remove noise from noisy latent vectors, and at sampling time the network is applied iteratively to Gaussian noise to produce a video.[23] The diffusion process runs jointly over the temporal audio latents and the spatio-temporal video latents.[23] The training data consisted of images, videos, audio, and associated annotations. Google says it captioned the data at different levels of detail using multiple Gemini models, filtered out unsafe captions and personally identifiable information, filtered videos for compliance, safety, and quality, and deduplicated the data semantically across sources.[15][23] The model was trained on Google's Tensor Processing Units using JAX and ML Pathways.[15] Google has not published parameter counts, layer counts, or other architectural details, and figures circulating in third-party blogs are not confirmed by Google. ## Technical capabilities ### Resolution and format In the Gemini API, Veo 3 and Veo 3 Fast generated clips of up to eight seconds at 24 frames per second, one video per request; options for 4, 6, or 8 seconds were added in October 2025.[22][33] At the Gemini API launch in July 2025, output was 720p in 16:9; the September 2025 update added 1080p (for 16:9 only) and 720p vertical 9:16 output.[22][32] In Flow, Google AI Ultra subscribers had 1080p generation from launch.[5] Veo 3.1 added 4K output, which Google's announcement described as upscaling, in the January 2026 update, available through Flow, the Gemini API, and Vertex AI.[35][75] The eight-second clip length is a practical constraint. Longer sequences can be built by chaining clips in Flow's Scenebuilder, or with the Extend (scene extension) feature introduced with Veo 3.1, which generates each new segment from the last second of the previous clip.[13][26] In the Gemini API, generation requests took between 11 seconds and six minutes, and generated videos were stored on Google's servers for two days.[22] ### Prompt understanding Veo 3 accepts text prompts and, in supported configurations, an input image. Prompts can combine descriptions of scenes, characters, actions, camera movements, lighting, dialogue, and sound in a single request, as in the examples on Google's model page.[3] The model card says Veo 3 "is able to faithfully follow simple and complex instructions," and in Google's human-rater evaluation it performed best on overall preference and prompt following.[15] Prompt adherence was a stated focus of the Veo 3.1 update.[13] ### Image-to-video Veo 3 supports image-to-video generation, in which a still image serves as the first frame and the model generates motion and audio from it. Image-to-video came to the Gemini API in paid preview on July 31, 2025, billed at the same rate as text-to-video.[32][74] In the Gemini app, a photo-to-video feature that turns photos into eight-second clips with sound rolled out to Google AI Pro and Ultra subscribers on July 10, 2025.[45] In 2025, Google's Gemini app video generation page stated: "For now, the ability to generate a video from a photo is not available in the European Economic Area, Switzerland, or the United Kingdom."[39] In the Gemini API, requests from the EU, the UK, Switzerland, and the Middle East and North Africa were limited to the "allow_adult" setting for generating people.[22] ### Camera control Veo 3 takes camera direction through natural language. Users can describe camera movements such as push-ins, tracking shots, pans, and dolly moves within the prompt, as in the prompts on Google's model page.[3] Flow adds a Camera Controls feature that gives "direct control over camera motion, angles and perspectives."[26] Because generation is stochastic, the same prompt run twice will not reproduce the same camera path exactly. ### Style and tone Veo 3 accepts style guidance within prompts, from film looks and color grading to animation styles such as stop motion or a painterly 2D look, as Google's sample prompts show.[3] The model card notes that Veo 3 can generate "a wide range of cinematic and visual styles."[15] Google's technical report adds that the model "has a bias for generating cinematic footage, with frequent camera cuts and dramatic camera angles."[23] ## Variants ### Veo 3 (original) The original Veo 3 model, released May 20, 2025, generated eight-second videos with synchronized audio from a text or image prompt.[1][22] It was billed per second of generated video. In the Gemini API it cost $0.75 per second with audio at its July 2025 launch, cut to $0.40 per second in September 2025.[2][20] It was available to Google AI Ultra subscribers through Flow and the Gemini app.[5] ### Veo 3 Fast Veo 3 Fast is a lighter-weight variant of Veo 3 designed for faster generation and lower cost. Google said it was twice as fast to generate as Veo 3 at launch in June 2025, with 720p output; 1080p followed in the Gemini API in September 2025.[20][44] In the Gemini API it cost $0.40 per second with audio, falling to $0.15 per second in September 2025.[20][32] Google pitched it at workloads such as programmatic advertising and rapid A/B testing.[32] YouTube used a custom version of Veo 3 Fast for free video generation in Shorts.[48] ### Veo 3.1 Veo 3.1 was released on October 15, 2025, as an update to Veo 3, together with a Veo 3.1 Fast variant.[13][53] Key additions: - **Reference images (Ingredients to Video)**: Up to three reference images of a character, object, or scene guide generation, which helps keep a character consistent across shots.[53] - **Frames to Video**: Users supply a first and a last frame, and the model generates the transition between them, with audio.[13][53] - **Scene extension**: New clips are generated from the final second of the previous clip, allowing sequences of a minute or more.[13][53] - **Editing in Flow**: Flow added an Insert tool for adding objects to a scene, with object removal announced as coming later.[13] The January 2026 update added native vertical output to Ingredients to Video, improved character consistency, and 1080p and 4K upscaling in Flow, the Gemini API, and Vertex AI.[12][35] ### Veo 3.1 Lite Veo 3.1 Lite, released on March 31, 2026, is the lowest-cost model in the Veo 3.1 family. It generates 4-, 6-, or 8-second clips with audio at 720p or 1080p in 16:9 or 9:16, without 4K output, and cost $0.05 per second at 720p and $0.08 at 1080p in the Gemini API.[34][37] Its Gemini API preview endpoint was shut down on October 22, 2026, but it remained available in Flow and was still listed in Google Cloud's Agent Platform pricing.[7][40][56] ## How much does Veo 3 cost? Veo 3 has been sold in two ways: bundled into Google's consumer AI subscriptions, where usage is metered by generation limits or Flow credits, and through pay-per-second APIs for developers and businesses. Prices below are those Google published at the dates shown; Google has changed them several times. ### Google AI Ultra and Google AI Pro At launch on May 20, 2025, Veo 3 was available only to Google AI Ultra subscribers in the United States.[4][5] Google AI Ultra cost $249.99 per month, with an offer of 50% off for the first three months for first-time subscribers.[5] In Flow, Ultra added 1080p video generation, advanced camera controls, and early access to Veo 3.[5] DataCamp, writing two days after launch, reported that Ultra users started with 12,500 Flow credits and that each Veo 3 generation used 150 credits.[17] Google AI Pro ($19.99 per month) initially included Flow with Veo 2 and 100 generations per month.[26] From May 30, 2025, Pro subscribers in the countries where Ultra was available received limited Veo 3 access in Flow and a 10-video trial in the Gemini app.[16] On June 9, 2025, Google gave Pro subscribers three Veo 3 Fast generations per day in the Gemini app; at that point Flow allocated 1,000 monthly AI credits to Pro and 12,500 to Ultra, and a Veo 3 Fast generation cost 20 credits.[44] In July 2025 MIT Technology Review reported that each Veo 3 generation cost a minimum of 20 credits and that additional credits could be bought at $25 per 2,500.[46] On August 18, 2025, Google doubled the monthly Flow credits for Ultra subscribers from 12,500 to 25,000, saying that Flow users had created 100 million videos since its launch in May.[55] In September 2025, Google's Flow credits help page listed a Veo 3 Quality generation at 100 credits and a Veo 3 Fast generation at 20 credits for non-Ultra subscribers and 0 credits for Ultra subscribers.[56] On May 19, 2026, Google introduced a second Google AI Ultra plan "starting at $100/month," which included video generation with Gemini Omni.[59] As of October 2026, the Flow help page listed monthly allocations of 1,000 credits for Pro subscribers, 10,000 for "Ultra $100" subscribers, and 25,000 for "Ultra $200" subscribers, with Veo 3.1 Lite, Fast, and Quality generations costing 10, 20, and 100 credits for non-Ultra subscribers and 5, 10, and 100 credits for Ultra subscribers.[56] ### Flow Flow is a filmmaking tool launched by Google alongside Veo 3 at Google I/O 2025.[26] Google described it as "the only AI filmmaking tool custom-designed for Google's most advanced models," naming Veo, Imagen, and Gemini; Flow combines Veo with the [Imagen 4](https://aiwiki.ai/wiki/imagen_4) image model and with Gemini for prompting.[26] It is the successor to VideoFX, the Google Labs video experiment launched in 2024.[26] No Film School described it as built for creatives working in film and video.[25] At launch Flow's features included Camera Controls, Scenebuilder (for editing and extending shots with continuous motion and consistent characters), Asset Management, and Flow TV, a showcase of Veo clips that displays the prompts used to make them.[26] Flow launched for Google AI Pro and Ultra subscribers in the United States.[26] The October 2025 update brought audio to all of Flow's existing features and added an Insert tool for adding elements to a scene.[13] Flow launched with short films by filmmakers Google had partnered with: Dave Clark's "Freelancers," about two estranged adopted brothers on similar quests; Henry Daubrez's "Electric Pink," about his own creative journey; and Junie Lau's "Dear Stranger," about the love between a grandmother and grandchild across parallel worlds.[26] Google said the films were made using Flow "along with other tools and techniques."[26] ### Vertex AI Vertex AI is [Google Cloud](https://aiwiki.ai/wiki/google_cloud)'s [enterprise AI](https://aiwiki.ai/wiki/enterprise_ai) platform. Veo 3 was in private preview on Vertex AI at launch, available to allowlisted accounts, and is billed per second of generated video.[1][7][57] Vertex AI customers can configure how aggressively prompts and outputs are filtered and control whether people can be generated, and, as on every Veo surface, outputs carry a SynthID watermark.[1] Veo 3 opened to all Google Cloud customers in public preview on June 27, 2025, and Veo 3 and Veo 3 Fast became generally available on Vertex AI on July 29, 2025.[57][60] Announcing general availability, Google said more than 70 million videos had been created with Veo 3 since May, and that enterprise customers had generated over 6 million videos on Vertex AI since its preview launch in June.[54] Unlike the Gemini API, Vertex AI priced video without audio separately. In July 2025 Veo 3 cost $0.75 per second with audio and $0.50 per second without; by October 2025 the prices were $0.40 and $0.20 for Veo 3, and $0.15 and $0.10 for Veo 3 Fast.[7] Veo 3.1 and Veo 3.1 Fast became generally available on Vertex AI on November 17, 2025.[57] On April 22, 2026, Vertex AI became part of Google's Gemini Enterprise Agent Platform; the Veo 3 and Veo 3 Fast models there carried a June 30, 2026 retirement date, although the platform's price list still included them alongside Veo 3.1, Veo 3.1 Fast, and Veo 3.1 Lite in October 2026.[7][57][58] ### Gemini API and Google AI Studio The Gemini API provides developer access to Veo 3 through Google AI Studio.[2] Veo 3 entered paid preview in the Gemini API on July 17, 2025, about two months after its announcement, and became generally available on September 9, 2025.[2][21][33] Access required a Google Cloud project on the paid tier, with no free tier for Veo generation.[32] Developers could start from a Veo 3 starter app in Google AI Studio, and Google cited early API users such as Cartwheel, which used Veo 3 to generate human motion for 3D animation, and the game studio Volley, which used it for in-game cut-scenes.[2] ### Price history All API prices are per second of generated video. From August 2025 the Gemini API listed only prices for video with audio; in July 2025 it also listed a video-only price that applied only when audio was blocked by safety filters. Vertex AI listed separate, lower prices for video without audio. | Date | Platform | Model | Price per second, with audio | Price per second, video only | |---|---|---|---|---| | July 2025 | Gemini API | Veo 3 | $0.75[2] | $0.50, charged only when audio was blocked by safety filters[34] | | July 2025 | Vertex AI | Veo 3 | $0.75[7] | $0.50[7] | | July 31, 2025 | Gemini API | Veo 3 Fast | $0.40[74] | n/a | | August 2025 | Vertex AI | Veo 3 Fast | $0.40[7] | $0.25[7] | | September 8-9, 2025 | Gemini API | Veo 3 / Veo 3 Fast | $0.40 / $0.15[20] | n/a | | October 2025 | Vertex AI | Veo 3 / Veo 3 Fast | $0.40 / $0.15[7] | $0.20 / $0.10[7] | | October 15, 2025 | Gemini API | Veo 3.1 / Veo 3.1 Fast | $0.40 / $0.15[34][53] | n/a | | February 2026 | Gemini API | Veo 3.1 / Veo 3.1 Fast at 4K | $0.60 / $0.35[34] | n/a | | March 31, 2026 | Gemini API | Veo 3.1 Lite | $0.05 (720p), $0.08 (1080p)[34][37] | n/a | | April 2026 | Gemini API | Veo 3.1 Fast (after April 7 cut) | $0.10 (720p), $0.12 (1080p), $0.30 (4K)[34][37] | n/a | | October 2026 | Agent Platform (formerly Vertex AI) | Veo 3.1 | $0.40 (720p and 1080p), $0.60 (4K)[7] | $0.20, $0.40 (4K)[7] | | October 2026 | Agent Platform (formerly Vertex AI) | Veo 3.1 Fast | $0.10 (720p), $0.12 (1080p), $0.30 (4K)[7] | $0.08, $0.10, $0.25 (4K)[7] | | October 2026 | Agent Platform (formerly Vertex AI) | Veo 3.1 Lite | $0.05 (720p), $0.08 (1080p)[7] | $0.03, $0.05[7] | | October 22, 2026 | Gemini API | All Veo 3.1 preview models | Shut down[40] | n/a | At the July 2025 Gemini API prices, an eight-second clip with audio cost $6.00 with Veo 3 or $3.20 with Veo 3 Fast.[32] At the September 2025 rates of $0.40 and $0.15 per second, the same clip cost $3.20 with Veo 3 or $1.20 with Veo 3 Fast.[20] ### YouTube Shorts YouTube CEO Neal Mohan announced at the Cannes Lions festival in June 2025 that Veo 3 would come to YouTube Shorts later that summer.[24] At its Made on YouTube event on September 16, 2025, YouTube launched a custom version of Veo 3 Fast in Shorts, free for creators, generating clips with sound at 480p with lower latency; it rolled out first in the United States, the United Kingdom, Canada, Australia, and New Zealand.[48] In January 2026 Veo 3.1's Ingredients to Video mode came to YouTube Shorts and the YouTube Create app, without the 1080p and 4K upscaling options reserved for Flow, the Gemini API, and Vertex AI.[12][35] ### Canva and other platforms In June 2025, [Canva](https://aiwiki.ai/wiki/canva) added Veo 3 to its platform through a "Create a Video Clip" feature that generates eight-second clips with sound from a text prompt.[11] It was available on Canva's paid plans (Pro, Teams, and Enterprise) and to Canva Nonprofit users, with an initial limit of five generations per month, and was also rolled out on [Leonardo.Ai](https://aiwiki.ai/wiki/leonardo_ai), which Canva owns.[11] Canva described itself as "one of the first platforms to integrate Google's Veo 3 video generation model, just weeks after its debut."[11] ## How does Veo 3 compare with other AI video models? When Veo 3 launched, most commercial video generators produced silent clips, and its native audio was the main point of comparison in coverage.[4][9] Over the following year several other developers released models that generate audio together with video, and by 2026 newer models, including Google's own Gemini Omni Flash, ranked above Veo on public preference leaderboards.[73] ### Feature comparison Capabilities as announced by each developer at release: | Model | Developer | Released | Native audio | Notes | |---|---|---|---|---| | Veo 3 | Google DeepMind | May 20, 2025 | Yes | 8-second clips; 720p, with 1080p added in the Gemini API in September 2025[20][22] | | Veo 3.1 | Google DeepMind | October 15, 2025 | Yes | Reference images, first and last frames, scene extension; 4K by upscaling from January 2026[12][53] | | Sora 2 | OpenAI | September 30, 2025 | Yes | Launched with an invite-based iOS app; discontinued in 2026[61][63] | | Kling 2.6 | Kuaishou | December 3, 2025 | Yes | Chinese and English speech; clips up to 10 seconds[64] | | Kling 3.0 | Kuaishou | February 2026 | Yes | Clips up to 15 seconds; speech in several languages; multi-character dialogue[65] | | Runway Gen-4 | Runway | March 31, 2025 | No | Consistent characters from a single reference image[66] | | Runway Gen-4.5 | Runway | December 1, 2025 | Added December 11, 2025 | Native audio added in an update released alongside Runway's GWM-1 world model[67][68] | | MiniMax Hailuo 02 | MiniMax | June 18, 2025 | Not announced | 768p clips of 6 or 10 seconds, or 1080p clips of 6 seconds[69] | | Pika 2.0 | Pika Labs | December 2024 | No | Pika had released a separate generative audio model for sound effects[70] | | Wan 2.1 | Alibaba | February 25, 2025 | No | Open weights under Apache 2.0; 480p and 720p[71] | | Wan 2.5-Preview | Alibaba | September 2025 | Yes | 10-second 1080p clips; not open source[72] | ### Leaderboard standings [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) runs a public text-to-video arena that ranks models by Elo scores computed from blind pairwise human preference votes.[73] When Runway released Gen-4.5 on December 1, 2025, it said the model led that leaderboard with 1,247 Elo points, ahead of Veo 3 at 1,226, Kling 2.5 at 1,225, and Sora 2 Pro at 1,206.[67] In archived snapshots of the leaderboard's ranking for video without audio, Kling 3.0 Pro was first on February 19, 2026, with Veo 3.1 fifth, and on March 22, 2026, Kling 3.0 Pro was still first while Veo 3.1 was twelfth.[73] In October 2026, on a rescaled board that fixes Kling 3.0 Pro at 1,000 points and ranks video generated with audio, Veo 3.1 was 22nd with 962 points, behind Google's own Gemini Omni Flash 1.1 in 10th place with 1,128; Alibaba's Wan 3.0 led with 1,175, and Veo 3 was no longer listed.[73] ### Veo 3 vs. Sora 2 OpenAI released Sora 2 on September 30, 2025, calling it its "flagship video and audio generation model," together with a new social iOS app called Sora.[61] OpenAI said Sora 2 could create "sophisticated background soundscapes, speech, and sound effects," putting it in direct competition with Veo 3 on audio.[61] In March 2026 OpenAI announced that it would discontinue Sora: the web and app versions closed on April 26, 2026, and the Sora API on September 24, 2026.[36][63] TechCrunch, citing The Wall Street Journal, reported that Sora's worldwide user count had peaked at around a million and then fallen below 500,000, that the app was "burning through roughly $1 million every day," and that Disney's $1 billion partnership with OpenAI ended with the shutdown.[62] ### Veo 3 vs. Runway [Runway](https://aiwiki.ai/wiki/runway_ml)'s Gen-4, released on March 31, 2025, focused on keeping characters, objects, and locations consistent across shots using reference images, and generated silent video.[66] Its successor, Gen-4.5, was released on December 1, 2025, and Runway added native audio to it on December 11, 2025, the same day it launched its first [world model](https://aiwiki.ai/wiki/world_model), GWM-1.[67][68] ### Veo 3 vs. Kling Kuaishou's [Kling](https://aiwiki.ai/wiki/kling) models were among Veo's closest competitors on preference leaderboards, and Google's own launch evaluation of Veo 3 included Kling 2.0 as a comparison model.[15][73] Kling gained what Kuaishou called "simultaneous audio-visual generation" with Kling 2.6 on December 3, 2025, supporting Chinese and English speech in clips of up to 10 seconds.[64] [Kling 3.0](https://aiwiki.ai/wiki/kling_3), announced on February 5, 2026, extended clips to 15 seconds and added native audio in several languages and accents, including multi-character dialogue scenes in which each character speaks a different language.[65] ### Veo 3 vs. Hailuo, Pika, and Wan MiniMax released Hailuo 02 on June 18, 2025, offering 768p clips of 6 or 10 seconds and 1080p clips of 6 seconds; its announcement did not mention audio.[69] [Pika Labs](https://aiwiki.ai/wiki/pika_labs) released Pika 2.0 in December 2024; the company had earlier released a generative audio model for sound effects.[70] Alibaba released the weights of Wan 2.1 under the Apache 2.0 license on February 25, 2025, with support for 480p and 720p video.[71] Its [Wan 2.5](https://aiwiki.ai/wiki/wan_2_5)-Preview, launched in September 2025, generates 10-second 1080p videos with audio tracks but, unlike earlier Wan models, is not open source.[72] ## Academic evaluation Veo 3 has been used as a test subject in research on whether video models can act as general-purpose vision systems. In "Video models are zero-shot learners and reasoners" (arXiv:2509.20328, September 2025), Google DeepMind researchers Thaddäus Wiedemer, Yuxuan Li, Paul Vicol, Shixiang Shane Gu, Nick Matarese, Kevin Swersky, Been Kim, Priyank Jaini, and Robert Geirhos prompted Veo 3 through Vertex AI with an input image and a text instruction, analyzing 18,384 generated videos across 62 qualitative and 7 quantitative tasks.[42] They reported that Veo 3 could solve tasks it was not trained for, such as segmenting objects, detecting edges, editing images, and understanding physical properties, and showed early forms of what they called "chain-of-frames" visual reasoning, such as solving mazes and symmetry puzzles.[42] On 5x5 mazes, Veo 3 reached a pass@10 rate of 78%, against 14% for Veo 2.[42] The authors noted that task-specific models still outperformed a zero-shot video model, and that because the Vertex AI API uses an LLM-based prompt rewriter, some solutions might come from the rewriter rather than the video model.[42] A follow-up study by Ziyu Guo and colleagues, "Are Video Models Ready as Zero-Shot Reasoners? An Empirical Study with the MME-CoF Benchmark" (arXiv:2510.26802, October 2025), evaluated Veo 3 across 12 reasoning dimensions. It concluded that current video models show promising short-horizon spatial coherence and locally consistent dynamics but "remain limited in long-horizon causal reasoning, strict geometric constraints, and abstract logic," and are "not yet reliable as standalone zero-shot reasoners."[43] ## What is Veo 3 used for? ### Independent filmmaking Veo 3 and Flow give independent filmmakers a way to produce short-form content without cameras, sets, or actors. Filmmakers use the model for establishing shots, b-roll, test renders of visual concepts, and rough character scenes with dialogue. The showcase filmmakers Google partnered with at Flow's launch (Dave Clark, Henry Daubrez, and Junie Lau) made narrative shorts rather than demo reels.[26] Director Darren Aronofsky used the tool to create a short film called "Ancestra," through his AI storytelling venture Primordial Soup, which announced a partnership with Google at I/O 2025.[25][46] Film-industry commentators have pointed to unresolved questions about [generative AI](https://aiwiki.ai/wiki/generative_ai) in production. Writing in January 2026 about the continuing partnership between Google DeepMind and Primordial Soup, CineD said "the legal and ethical frameworks around synthetic performance and character creation remain unsettled."[35] ### Advertising and marketing Marketing teams use Veo 3 to generate concept videos, product clips, and social media content, and to test several creative directions before committing production budgets. In the Gemini API, Google pitched Veo 3 Fast at uses such as programmatic advertising and rapid A/B testing, and early users made mock pharmaceutical ads and other commercials with the model.[9][32] MIT Technology Review quoted an advertising creative director who said that up to 40% of her dialogue scenes came out with gibberish subtitles, making them unusable and expensive to regenerate.[46] ### Social media content Creators on platforms such as [YouTube](https://aiwiki.ai/wiki/youtube), [TikTok](https://aiwiki.ai/wiki/tiktok), and Instagram use Veo 3 to generate short clips with audio, either as finished posts or as parts of longer edits. Common formats in the months after launch included staged man-on-the-street interviews, ASMR clips, fake product reviews, and fictional news segments.[9][14][45][46] Africa Check described Veo 3 clips on TikTok depicting African settings, including market scenes and dialogue in Nigerian Pidgin, some of which crossed into deception with fabricated weather events, bogus health product reviews, and staged news broadcasts.[14] YouTube's free Veo 3 Fast integration in Shorts, launched in September 2025, put the model directly into a major short-video app.[48] ### Prototyping and storyboarding Production teams use Veo 3 Fast and later Veo 3.1 to generate animated storyboards and rough cuts before entering full production. The eight-second clip length matches the shot-level granularity of storyboarding. Google cited Promise Studios, which used Veo 3.1 in its MUSE platform for generative storyboarding and previsualization.[53] ### Games, animation, and software Developers have built Veo 3 into other products through the Gemini API. Google cited Cartwheel, which turns generated human motion into 3D animation on rigged characters, and Volley, which used Veo 3 for cut-scenes in its role-playing game "Wit's End."[2] Latitude experimented with Veo 3.1 in its narrative game engine.[53] ## SynthID watermarking All videos generated by Veo 3 carry an invisible [SynthID](https://aiwiki.ai/wiki/synthid) watermark embedded by Google DeepMind.[1][16] SynthID embeds a signal that is imperceptible to viewers but can be read by Google's detection tools.[1] Google describes SynthID watermarking as one of Veo 3's post-training mitigations against misinformation.[15][23] On May 30, 2025, Google also began adding a visible watermark to all Veo 3 videos except those generated by Ultra subscribers in Flow.[16] The mark, a small "Veo" label in the lower right corner, was criticized as easy to miss and easy to crop out.[9][10][14] TIME reported that Google decided to add it after TIME contacted the company about deepfakes it had generated with the model.[9] Videos without the visible mark still carry SynthID.[10] ### Detection At launch, Google's SynthID Detector was available only to early testers, and journalists and fact-checkers criticized the lack of a public way to check whether a video came from Veo 3.[9][10][14] On December 18, 2025, Google added video verification to the Gemini app: users can upload a video of up to 100 MB and 90 seconds and ask whether it was made with Google AI, and Gemini checks the audio and visual tracks for SynthID and reports which segments contain it.[49] The feature is available in all languages and countries where the Gemini app is offered.[49] ### Limitations of SynthID SynthID only identifies content generated by Google AI tools, so videos from Sora, Runway, Kling, or other non-Google systems are outside its scope.[49] Academic work has shown that invisible watermarks can be attacked. UnMarker, an attack on image watermarking by Andre Kassis and Urs Hengartner of the University of Waterloo, presented at the 2025 IEEE Symposium on Security and Privacy, removed watermarks more than 50% of the time across the schemes tested, including Google's SynthID, without knowledge of the watermarking method.[50] That research concerned still images rather than video. In practice, SynthID works as a check for platforms, fact-checkers, and users who think to run it, while the visible watermark gives a simpler signal that can be cropped out.[9][49] ## Safety and responsible AI Veo 3's model card, first published on May 23, 2025 and updated on January 13, 2026 to cover subsequent versions, describes the safety evaluation that preceded launch.[15] Development evaluations were designed internally; assurance evaluations were developed and run "by a group outside of the model development team"; red teaming was done by a mix of specialist internal and external teams; and ethics and safety reviews were conducted before release.[15] Google's technical report says that before launch, the Google DeepMind Responsibility and Safety Council reviews a model's performance across these evaluations to make release decisions.[23] Mitigations included safety filtering of pre-training data, removal of duplicated and conceptually similar videos, synthetic captions to improve the variety of concepts in the training data, and, after training, SynthID watermarking and production filtering.[15] Use of the model is governed by Google's Generative AI Prohibited Use Policy, which bars uses including dangerous or illegal activities, compromising security, sexually explicit, violent, hateful, or harmful content, and misinformation or misleading activities.[15] The model card reports several findings. During testing for unfair bias, Google "noted that Veo 3 appears to skew towards lighter skin tones when race is not specified in the prompt," and testing "surfaced risks of semantic bias where particular terms are spuriously correlated with representation of particular demographics."[15] These findings were shared with model and product teams, though the card gives no timelines or quantitative targets.[15] Google found little evidence of risk around self-replication, tool use, or cybersecurity, and limited capability in chemical, biological, radiological, nuclear, and explosives domains.[15] It acknowledged that Veo 3 "was found to be able to produce deepfakes," though it said these were of worse quality than those made by dedicated deepfake tools.[15] The technical report adds that the image-to-video model could produce "quite good deep-fakes" that were still less controllable than dedicated tools, "particularly with respect to speech."[23] ## Reception Veo 3 drew wide attention after its May 2025 launch.[4] The synchronized audio produced viral social media moments, as creators shared clips with realistic dialogue, ambient sound, and music generated from text prompts, including short films, mock pharmaceutical ads, and man-on-the-street interviews.[9] Within seven weeks, Google said users had generated more than 40 million Veo 3 videos in the Gemini app and Flow.[45] Technology press coverage was largely positive about the technical advance. CNBC said the ability to incorporate audio was "a key distinction" from OpenAI's Sora.[4] TIME called Veo 3 "a significant jump forward, creating AI clips that are nearly indistinguishable from real ones."[9] DataCamp's Alex Olteanu wrote that "Veo 3 delivers something that feels fundamentally new: coherent, sound-enabled video from natural language prompts," while noting prompt drift, occasional visual glitches, and the slow and costly iteration that the credit system imposed.[17] Some commentators were more critical of the timing. Katerina Cizek of the MIT Open Documentary Lab told MIT Technology Review that "Google needed a win" and needed "to be the first to pump out a tool that generates lip-synched audio," which she said mattered more to the company than fixing the model's subtitle problem.[46] ### The Will Smith spaghetti test The "Will Smith eating spaghetti" test became one of the most discussed moments of Veo 3's launch. The benchmark began on March 27, 2023, when a Reddit user posted an AI-generated clip made with [ModelScope](https://aiwiki.ai/wiki/modelscope)'s text-to-video tool that showed a grotesquely distorted Will Smith eating spaghetti.[19] The clip went viral, and "eating spaghetti" became shorthand for testing how far a new video model had progressed on realistic human motion and eating.[18][19] On May 22, 2025, two days after Veo 3's launch, AI content creator Javi Lopez posted his Veo 3 version of the test on X.[18][31] The video showed a recognizable Will Smith eating and slurping noodles with generated audio.[18] PetaPixel headlined that Veo 3 "nails" the test.[18] The audio drew most of the comment: Smith made crunching and squelching noises while chewing, which led YouTuber Marques Brownlee to comment "I don't feel so good," and one X user asked, "Why is his spaghetti crunchy?"[18][31] The test illustrated a broader pattern in Veo 3's reception: the model was clearly a step change over earlier tools, but its realism exposed new ways for output to be subtly wrong, with audio mistakes and small physical errors becoming the new focus of critique. ### Misinformation concerns Media organizations and researchers raised concerns about the model's potential for misinformation. In a June 3, 2025 investigation, TIME reported that it had used Veo 3 to create realistic videos of a Pakistani crowd setting fire to a Hindu temple, Chinese researchers handling a bat in a wet lab, an election worker shredding ballots, and Palestinians gratefully accepting U.S. aid in Gaza.[9] Experts told TIME that while each video had noticeable inaccuracies, such clips could fuel unrest or violence if shared with a misleading caption during breaking news.[9] TIME also reported that the model refused some prompts, generally declining to depict recognizable public figures such as Donald Trump and Elon Musk.[9] [Connor Leahy](https://aiwiki.ai/wiki/connor_leahy), CEO of the [AI safety](https://aiwiki.ai/wiki/ai_safety) company [Conjecture](https://aiwiki.ai/wiki/conjecture), told TIME that "The risks from deepfakes and synthetic media have been well known and obvious for years, and the fact the tech industry can't even protect against such well-understood, obvious risks is a clear warning sign that they are not responsible enough to handle even more dangerous, uncontrolled AI and AGI."[9] Julia Smakman, a researcher at the Ada Lovelace Institute, said in the same report that existing technical safeguards such as safety classifiers were "proving insufficient to stop harmful images and videos from being generated."[9] TIME also reported that in the first week of Veo 3's release, users posted fake news segments in multiple languages, including an anchor announcing the death of J.K. Rowling and fake political news conferences.[9] The fact-checking organization Africa Check published a guide to identifying Veo 3 videos, pointing to the small visible watermark, garbled on-screen text, mispronounced local-language words, scenes that never run longer than about eight seconds, and glitches such as objects suddenly appearing or disappearing.[14] In July 2025, Media Matters for America reported that racist and antisemitic videos apparently made with Veo 3 were drawing millions of views on TikTok; the clips ran no longer than eight seconds and many carried the "Veo" watermark.[51] ### Training data and copyright Google has not said exactly what data Veo 3 was trained on. In June 2025 CNBC reported that Google was using its library of YouTube videos to train AI models including Gemini and Veo 3; Google confirmed it uses a subset of YouTube videos for training and said it honors specific agreements with creators and media companies.[47] Creators and intellectual-property professionals interviewed by CNBC said they had not been told their content could be used this way.[47] In December 2025, Disney sent Google a cease-and-desist letter accusing it of copyright infringement on a "massive scale," naming Veo along with Imagen, [Nano Banana](https://aiwiki.ai/wiki/nano_banana), Gemini, and YouTube, and citing characters from franchises such as "Frozen," "The Lion King," "Moana," "Deadpool," and "Star Wars."[52] Disney demanded that Google stop distributing infringing outputs and stop training on its works.[52] ### Industry response After Veo 3's launch, several other developers released video models that generate audio together with the picture: Alibaba's Wan 2.5-Preview and OpenAI's Sora 2 in September 2025, Kuaishou's Kling 2.6 in December 2025, and Runway's Gen-4.5, to which native audio was added in December 2025.[61][64][68][72] Kuaishou described its Kling 2.6 release as reshaping a workflow that "typically requires generating silent footage first and using separate software for post-production audio."[64] ## What are the limitations of Veo 3? Veo 3 has several documented limitations: **Clip length**: Each generation is limited to eight seconds. Longer content requires stitching clips together in Flow or using the scene extension capability introduced with Veo 3.1.[13][22] **Complex scenes**: Google's model card says that "maintaining complete consistency throughout complex scenes or those with complex motion, remains a challenge."[15] Characters and objects can change appearance between shots; the Ingredients to Video feature in Veo 3.1 was designed to reduce this for characters.[13][53] **Text rendering**: On-screen text such as signs, captions, and labels often comes out garbled. Google's technical report says Veo 3 "is still poor at generating text," and Africa Check found illegible headlines and misspelled signs in Veo 3 news-style clips.[14][23] **Unwanted subtitles**: Veo 3 often added nonsensical subtitles to clips with dialogue, even when prompts asked for none. Google said on June 9, 2025 that it had developed fixes, but MIT Technology Review reported in July 2025 that users were still encountering the problem; experts attributed it to subtitled videos in the training data.[44][46] **Audio errors**: Generated audio can mismatch the action, as in the crunching spaghetti of the Will Smith test, and pronunciation of less common languages can be wrong.[14][18] **Regional restrictions**: Some features rolled out region by region. Photo-to-video in the Gemini app excluded the European Economic Area, Switzerland, and the United Kingdom, and in the Gemini API, generation of people was restricted in the EU, the UK, Switzerland, and the Middle East and North Africa.[22][39] **Access cost**: At launch, full Veo 3 access required the Google AI Ultra plan at $249.99 per month.[4][5] Veo 3 Fast, the September 2025 API price cuts, and free Veo 3 Fast generation in YouTube Shorts made the model progressively cheaper to use.[20][44][48] **Misinformation risk**: The realism of Veo 3's audio-visual output, particularly dialogue with lip-sync, makes it easier to produce [deepfake](https://aiwiki.ai/wiki/deepfake) and other fabricated videos that are hard to recognize as AI-generated.[9] This has been documented by journalists and researchers.[9][14][51] ## See also - [Veo](https://aiwiki.ai/wiki/veo) - [Sora](https://aiwiki.ai/wiki/sora) - [Sora 2](https://aiwiki.ai/wiki/sora_2) - [Kling](https://aiwiki.ai/wiki/kling) - [Runway ML](https://aiwiki.ai/wiki/runway_ml) - [Pika Labs](https://aiwiki.ai/wiki/pika_labs) - [Hailuo AI](https://aiwiki.ai/wiki/hailuo) - [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind) - [Vertex AI](https://aiwiki.ai/wiki/vertex_ai) - [Diffusion model](https://aiwiki.ai/wiki/diffusion_model) - [SynthID](https://aiwiki.ai/wiki/synthid) - [Imagen 4](https://aiwiki.ai/wiki/imagen_4) - [Gemini](https://aiwiki.ai/wiki/gemini) - [Gemini Omni](https://aiwiki.ai/wiki/gemini_omni) - Flow (Google) - [Generative AI](https://aiwiki.ai/wiki/generative_ai) - [Text-to-video generation](https://aiwiki.ai/wiki/text_to_video) ## References 1. Google Cloud. "Announcing Veo 3, Imagen 4, and Lyria 2 on Vertex AI" (published as "Expanding Vertex AI with the next wave of generative AI media models"). Google Cloud Blog, May 21, 2025. https://cloud.google.com/blog/products/ai-machine-learning/announcing-veo-3-imagen-4-and-lyria-2-on-vertex-ai 2. Google Developers. "Build with Veo 3, now available in the Gemini API." Google Developers Blog, July 17, 2025. https://developers.googleblog.com/en/veo-3-now-available-gemini-api/ 3. Google DeepMind. "Veo." Google DeepMind models page (accessed October 30, 2026). https://deepmind.google/models/veo/ 4. CNBC. "Google launches Veo 3, an AI video generator that incorporates audio." May 20, 2025. https://www.cnbc.com/2025/05/20/google-ai-video-generator-audio-veo-3.html 5. Google. "Introducing Google AI Ultra: The best of Google AI in one subscription." The Keyword, May 20, 2025. https://blog.google/products-and-platforms/products/google-one/google-ai-ultra/ 6. Google Flow Help. "Learn about Google Flow models & supported features" (accessed October 30, 2026). https://support.google.com/flow/answer/16352836 . Flow itself moved from labs.google/flow to https://flow.google.com/ 7. Google Cloud. "Vertex AI pricing" (now "Agent Platform pricing," accessed October 30, 2026). https://cloud.google.com/vertex-ai/generative-ai/pricing ; archived snapshots: July 5, 2025 https://web.archive.org/web/20250705053458/https://cloud.google.com/vertex-ai/generative-ai/pricing ; August 6, 2025 https://web.archive.org/web/20250806164725/https://cloud.google.com/vertex-ai/generative-ai/pricing ; October 2, 2025 https://web.archive.org/web/20251002151121/https://cloud.google.com/vertex-ai/generative-ai/pricing ; February 6, 2026 https://web.archive.org/web/20260206230133/https://cloud.google.com/vertex-ai/generative-ai/pricing 8. Google AI for Developers. "Video generation in the Gemini API" (accessed October 30, 2026; notes the October 22, 2026 shutdown of Veo 3.1 Preview). https://ai.google.dev/gemini-api/docs/video 9. Andrew R. Chow and Billy Perrigo. "Google's New AI Tool Generates Convincing Deepfakes of Riots, Conflict, and Election Fraud." TIME, June 3, 2025. https://time.com/7290050/veo-3-google-misinformation-deepfake/ 10. BGR. "Those Amazing Veo 3 Videos Will Finally Tell You They Were Made With AI." June 2, 2025. https://www.bgr.com/tech/those-amazing-veo-3-videos-will-finally-tell-you-they-were-made-with-ai/ 11. Canva. "Google's Veo 3 comes to Canva: Introducing Create a Video Clip." Canva Newsroom, June 2025. https://www.canva.com/newsroom/news/veo3-canva-ai-video/ 12. Google. "Veo 3.1 Ingredients to Video: More consistency, creativity and control." The Keyword, January 13, 2026. https://blog.google/innovation-and-ai/technology/ai/veo-3-1-ingredients-to-video/ 13. Google. "Introducing Veo 3.1 and advanced capabilities in Flow." The Keyword, October 15, 2025. https://blog.google/innovation-and-ai/products/veo-updates-flow/ 14. Africa Check. "GUIDE: How to tell if a video was made with Google's new 'state-of-the-art' Veo 3." 2025. https://africacheck.org/fact-checks/guides/guide-how-tell-if-video-was-made-googles-new-state-art-veo-3 15. Google DeepMind. "Veo 3 Model Card." Published May 23, 2025; updated January 13, 2026. https://storage.googleapis.com/deepmind-media/Model-Cards/Veo-3-Model-Card.pdf 16. Google. "We're bringing Veo 3 to more countries, and to more users on the Gemini mobile app." The Keyword, May 30, 2025. https://blog.google/products-and-platforms/products/gemini/veo-3-expansion-mobile/ 17. Alex Olteanu. "Google's Veo 3: A Guide With Practical Examples." DataCamp, May 22, 2025. https://www.datacamp.com/tutorial/veo-3 18. PetaPixel. "Google's Veo 3 Nails the Infamous Will Smith Eating Spaghetti Test." May 28, 2025. https://petapixel.com/2025/05/28/googles-veo-3-nails-the-infamous-will-smith-eating-spaghetti-test/ 19. Wikipedia. "Will Smith eating spaghetti test." https://en.wikipedia.org/wiki/Will_Smith_eating_spaghetti_test 20. Google Developers. "Veo 3 and Veo 3 Fast: new pricing, new configurations and better resolution." Google Developers Blog, September 8, 2025. https://developers.googleblog.com/en/veo-3-and-veo-3-fast-new-pricing-new-configurations-and-better-resolution/ 21. Neowin. "Google Veo 3 and Veo 3 Fast models get a price cut and new video features." September 9, 2025. https://www.neowin.net/news/google-veo-3-and-veo-3-fast-models-get-a-price-cut-and-new-video-features/ 22. Google AI for Developers. "Generate videos with Veo 3 in Gemini API." Archived September 23, 2025. https://web.archive.org/web/20250923203433/https://ai.google.dev/gemini-api/docs/video 23. Google DeepMind. "Veo 3 Technical Report." https://storage.googleapis.com/deepmind-media/veo/Veo-3-Tech-Report.pdf 24. The Hollywood Reporter. "YouTube to Add Google's Veo 3 to Shorts in Move That Could Turbocharge AI on the Video Platform." June 18, 2025. https://www.hollywoodreporter.com/business/digital/youtube-add-google-veo-3-shorts-ai-1236293135/ 25. No Film School. "Google Introduces Flow: A New AI Filmmaking Tool Powered by Veo 3." May 21, 2025. https://nofilmschool.com/google-flow-ai-filmmaking-tool 26. Google. "Meet Flow: AI-powered filmmaking with Veo 3." The Keyword, May 20, 2025. https://blog.google/innovation-and-ai/products/google-flow-veo-ai-filmmaking-tool/ 27. Google. "State-of-the-art video and image generation with Veo 2 and Imagen 3." The Keyword, December 16, 2024. https://blog.google/innovation-and-ai/models-and-research/google-labs/video-image-generation-update-december-2024/ 28. Google. "Google I/O 2025: Sundar Pichai's opening keynote." The Keyword, May 20, 2025. https://blog.google/innovation-and-ai/technology/ai/io-2025-keynote/ 29. 9to5Google. "Gemini app 'Pro' users getting Veo 3 trial on Android & iOS, AI Ultra now in 73 countries." May 30, 2025. https://9to5google.com/2025/05/30/gemini-app-veo-3-ai-ultra/ 30. Wikipedia. "Veo (text-to-video model)." https://en.wikipedia.org/wiki/Veo_(text-to-video_model) 31. The Daily Dot. "'Super al dente': The Will Smith eating spaghetti meme gets a high-resolution upgrade with Google's Veo 3." 2025. https://www.dailydot.com/news/will-smith-ai-spaghetti-google-veo/ 32. Matthias Bastian. "Google Veo 3 adds 1080p, 9:16 video, and drops prices by half." The Decoder, updated September 9, 2025. https://the-decoder.com/google-veo-3-adds-1080p-916-video-and-drops-prices-by-half/ 33. Google AI for Developers. "Release notes" (Gemini API changelog, accessed October 30, 2026). https://ai.google.dev/gemini-api/docs/changelog 34. Google AI for Developers. "Gemini Developer API pricing." https://ai.google.dev/gemini-api/docs/pricing ; archived snapshots: July 25, 2025 https://web.archive.org/web/20250725/https://ai.google.dev/gemini-api/docs/pricing ; November 3, 2025 https://web.archive.org/web/20251103074340/https://ai.google.dev/gemini-api/docs/pricing ; February 1, 2026 https://web.archive.org/web/20260201102619/https://ai.google.dev/gemini-api/docs/pricing ; April 1, 2026 https://web.archive.org/web/20260401191258/https://ai.google.dev/gemini-api/docs/pricing ; May 1, 2026 https://web.archive.org/web/20260501160638/https://ai.google.dev/gemini-api/docs/pricing 35. CineD. "Google Veo 3.1 Ingredients to Video Update Adds Native Vertical Format, 4K Upscaling, and Enhanced Character Consistency." January 2026. https://www.cined.com/google-veo-3-1-ingredients-to-video-update-adds-native-vertical-format-4k-upscaling-and-enhanced-character-consistency/ 36. OpenAI Help Center. "What to know about the Sora discontinuation." 2026. https://help.openai.com/en/articles/20001152-what-to-know-about-the-sora-discontinuation 37. Google. "Build with Veo 3.1 Lite, our most cost-effective video generation model." The Keyword, March 31, 2026. https://blog.google/innovation-and-ai/technology/ai/veo-3-1-lite/ 38. Koray Kavukcuoglu. "Introducing Gemini Omni." The Keyword, May 19, 2026. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni/ 39. Google. Gemini app video generation overview (accessed October 30, 2026). https://gemini.google/overview/video-generation/ ; archived September 1, 2025 as "Gemini AI video generator powered by Veo 3": https://web.archive.org/web/20250901183455/https://gemini.google/overview/video-generation/ 40. Google AI for Developers. "Gemini deprecations" (accessed October 30, 2026). https://ai.google.dev/gemini-api/docs/deprecations 41. TechCrunch. "Google's Veo 3 can generate videos, and soundtracks to go along with them." May 20, 2025. https://techcrunch.com/2025/05/20/googles-veo-3-can-generate-videos-and-soundtracks-to-go-along-with-them/ 42. Thaddäus Wiedemer, Yuxuan Li, Paul Vicol, Shixiang Shane Gu, Nick Matarese, Kevin Swersky, Been Kim, Priyank Jaini, and Robert Geirhos. "Video models are zero-shot learners and reasoners." arXiv:2509.20328, September 2025. https://arxiv.org/abs/2509.20328 43. Ziyu Guo et al. "Are Video Models Ready as Zero-Shot Reasoners? An Empirical Study with the MME-CoF Benchmark." arXiv:2510.26802, October 2025. https://arxiv.org/abs/2510.26802 44. 9to5Google. "Gemini app Pro subscribers now get three 'Veo 3 Fast' videos per day." June 9, 2025. https://9to5google.com/2025/06/09/gemini-veo-3-fast/ 45. Google. "Turn your photos into videos in Gemini." The Keyword, July 10, 2025. https://blog.google/products-and-platforms/products/gemini/photo-to-video/ 46. MIT Technology Review. "Google's generative video model Veo 3 has a subtitles problem." July 15, 2025. https://www.technologyreview.com/2025/07/15/1120156/googles-generative-video-model-veo-3-has-a-subtitles-problem/ 47. CNBC. "Creators say they didn't know Google uses YouTube to train AI." June 19, 2025. https://www.cnbc.com/2025/06/19/google-youtube-ai-training-veo-3.html 48. YouTube Official Blog. "Unpacking the magic of our new creative tools." September 16, 2025. https://blog.youtube/news-and-events/generative-ai-creation-tools-made-on-youtube-2025/ 49. Google. "You can now verify Google AI-generated videos in the Gemini app." The Keyword, December 18, 2025. https://blog.google/innovation-and-ai/products/verify-google-ai-videos-gemini-app/ 50. University of Waterloo. "Watermarks offer no defense against deepfakes." Waterloo News, July 23, 2025. https://uwaterloo.ca/news/media/watermarks-offer-no-defense-against-deepfakes 51. Media Matters for America. "Racist AI-generated videos are the newest slop garnering millions of views on TikTok." July 2025. https://www.mediamatters.org/tiktok/racist-ai-generated-videos-are-newest-slop-garnering-millions-views-tiktok 52. TheWrap. "Disney Sends Cease and Desist Letter to Google Alleging Copyright Infringement on 'Massive Scale' in Gen AI Results." December 11, 2025. https://www.thewrap.com/disney-google-cease-and-desist-letter-copyright-infringement-allegations/ 53. Google Developers. "Introducing Veo 3.1 and new creative capabilities in the Gemini API." Google Developers Blog, October 15, 2025. https://developers.googleblog.com/introducing-veo-3-1-and-new-creative-capabilities-in-the-gemini-api/ 54. Jason Gelman. "Veo 3 and Veo 3 Fast are now generally available on Vertex AI." Google Cloud Blog, July 30, 2025. https://cloud.google.com/blog/products/ai-machine-learning/veo-3-fast-available-for-everyone-on-vertex-ai 55. Niv Govindaraju. "Get more out of Flow in Google AI Ultra." The Keyword, August 18, 2025. https://blog.google/innovation-and-ai/models-and-research/google-labs/flow-whisk-ai-credits-update/ 56. Google Flow Help. "Manage your Google Flow credits" (accessed October 30, 2026). https://support.google.com/flow/answer/16526234 ; earlier version archived September 19, 2025: https://web.archive.org/web/20250919065824/https://support.google.com/labs/answer/16526234 57. Google Cloud. "Vertex AI release notes" and "Gemini Enterprise Agent Platform release notes" (accessed October 30, 2026). https://docs.cloud.google.com/vertex-ai/generative-ai/docs/release-notes ; https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes 58. Google Cloud. "Model versions and lifecycle." Gemini Enterprise Agent Platform documentation (accessed October 30, 2026). https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-versions 59. Google. "Gemini Apps' release updates & improvements" (entry of May 19, 2026). https://gemini.google/release-notes/ 60. Jason Gelman. "You dream it, Veo creates it: Veo 3 is now available for everyone in public preview on Vertex AI." Google Cloud Blog, June 27, 2025. https://cloud.google.com/blog/products/ai-machine-learning/veo-3-available-for-everyone-in-public-preview-on-vertex-ai 61. OpenAI. "Sora 2 is here." September 30, 2025. https://openai.com/index/sora-2/ 62. TechCrunch. "Why OpenAI really shut down Sora." March 29, 2026. https://techcrunch.com/2026/03/29/why-openai-really-shut-down-sora/ 63. Matthias Bastian. "OpenAI sets two-stage Sora shutdown with app closing April 2026 and API following in September." The Decoder, March 28, 2026. https://the-decoder.com/openai-sets-two-stage-sora-shutdown-with-app-closing-april-2026-and-api-following-in-september/ 64. Kuaishou Technology. "Kling AI Launches Video 2.6 Model with 'Simultaneous Audio-Visual Generation' Capability, Redefining AI Video Creation Workflow." December 5, 2025. https://kuaishou.gcs-web.com/news-releases/news-release-details/kling-ai-launches-video-26-model-simultaneous-audio-visual 65. Kuaishou Technology. "Kling AI Launches 3.0 Model, Ushering in an Era Where Everyone Can Be a Director." February 5, 2026. https://kuaishou.gcs-web.com/news-releases/news-release-details/kling-ai-launches-30-model-ushering-era-where-everyone-can-be 66. Runway. "Introducing Runway Gen-4." March 31, 2025. https://runwayml.com/research/introducing-runway-gen-4 67. Runway. "Introducing Runway Gen-4.5." December 1, 2025. https://runwayml.com/research/introducing-runway-gen-4.5 68. Ivan Mehta and Rebecca Bellan. "Runway releases its first world model, adds native audio to latest video model." TechCrunch, December 11, 2025. https://techcrunch.com/2025/12/11/runway-releases-its-first-world-model-adds-native-audio-to-latest-video-model/ 69. MiniMax. "MiniMax Hailuo 02, World-Class Quality, Record-Breaking Cost Efficiency." June 18, 2025. https://www.minimax.io/news/minimax-hailuo-02 70. Matthias Bastian. "Pika Labs releases AI video generator 2.0 with new features." The Decoder, December 14, 2024. https://the-decoder.com/pika-labs-releases-ai-video-generator-2-0-with-new-features/ 71. Wan-Video. "Wan2.1" (GitHub repository README). https://github.com/Wan-Video/Wan2.1 72. Jonathan Kemper. "Alibabas Wan2.5-Preview lets users turn photos and text prompts into videos with matching audio." The Decoder, September 25, 2025. https://the-decoder.com/alibabas-wan2-5-preview-lets-users-turn-photos-and-text-prompts-into-videos-with-matching-audio/ 73. Artificial Analysis. "Text to Video Leaderboard" (AA-Video-T2V v2.1, accessed October 30, 2026). https://artificialanalysis.ai/video/leaderboard/text-to-video ; archived snapshots: February 19, 2026 https://web.archive.org/web/20260219141621/https://artificialanalysis.ai/video/leaderboard/text-to-video ; March 22, 2026 https://web.archive.org/web/20260322232610/https://artificialanalysis.ai/video/leaderboard/text-to-video 74. Google Developers. "Veo 3 Fast and new image-to-video capabilities." Google Developers Blog, July 31, 2025. https://developers.googleblog.com/en/veo-3-fast-image-to-video-capabilities-now-available-gemini-api/ 75. Google AI for Developers. "Generate videos with Veo 3.1 in Gemini API." Archived February 1, 2026. https://web.archive.org/web/20260201/https://ai.google.dev/gemini-api/docs/video