Veo 2

Veo 2 is a text-to-video generative artificial intelligence model developed by Google DeepMind and announced on December 16, 2024, the second major iteration of the Veo family. It generates short clips from text prompts or a starting image. At launch Google said it would deliver video "at resolutions up to 4K, and extended to minutes in length," with an improved understanding of real-world physics, more natural human motion, and less frequent hallucination of unwanted details such as extra fingers.^[1] In practice, the Google products and APIs that documented their Veo 2 output produced 720p clips of up to eight seconds per generation: VideoFX capped videos at 720p and eight seconds, and the Gemini API and Vertex AI generated 720p, 24 fps clips of five to eight seconds.^[8][4][5] TechCrunch described it as a model "to rival Sora," and it arrived one week after OpenAI released Sora Turbo to ChatGPT Plus and Pro subscribers.^[8][7] In Google's own human-rater comparison on 1,003 MovieGenBench prompts, raters preferred Veo 2 over Sora Turbo 58.8% of the time and Sora Turbo 26.7% of the time, with 14.5% ties.^[2][19]

The model arrived first as a waitlisted preview inside VideoFX in Google Labs, then expanded in stages. Veo 2 came to YouTube Shorts on February 13, 2025 (in the United States, Canada, Australia and New Zealand). It became generally available in the Gemini API on April 9, 2025, and reached Gemini Advanced subscribers in the Gemini app and the Whisk Animate feature on April 15, 2025. It came to Google Vids on May 6, 2025.^{[9][4][3][24]} On Vertex AI, Veo was an allowlisted experimental offering in early 2025, and veo-2.0-generate-001 was listed as "GA allowlist" by April 2025. Google's lifecycle table gives veo-2.0-generate-001 a release date of May 27, 2025.^[5][32] Veo 3 followed on May 20, 2025 and Veo 3.1 on October 15, 2025, and Veo 2 stayed on sale alongside them. Google retired Veo 2 on Vertex AI and shut it down in the Gemini API on June 30, 2026. On Vertex AI it recommended veo-3.1-generate-001 as the upgrade. In the Gemini API it listed gemini-omni-1.1-flash or the generally available models on the Gemini Enterprise Agent Platform as replacements.^[6][32][33]

Infobox

What is Veo 2?

Veo 2 is the December 2024 text-to-video and image-to-video model from Google DeepMind. It succeeded the original Veo model of May 2024 and preceded Veo 3. It accepts a text prompt, or an image used as the first frame, and outputs a short MP4 clip.^[1][35] Google's headline claims over the first Veo were a higher maximum resolution (4K, against 1080p for the first Veo), a better grasp of real-world physics and human movement, and stronger interpretation of cinematic language.^[1][44] Like the first Veo, it embeds an invisible SynthID watermark in its output.^[1][18] In Google DeepMind's framing, Veo 2 "creates incredibly high-quality videos in a wide range of subjects and styles" and "achieved state-of-the-art results against leading models" when judged by human raters.^[1]

Background

Veo 1 and the May 2024 launch

Google announced the original Veo model, sometimes called Veo 1, at Google I/O on May 14, 2024. Google said it generated "high-quality 1080p resolution videos in a wide range of cinematic and visual styles that can go beyond a minute," and that it understood cinematic terms such as "timelapse" and "aerial shots of a landscape."^[44] It arrived while OpenAI's Sora, which OpenAI had released in beta to a select group of testers in February 2024, was not yet publicly available.^[30] Veo was opened to "select creators in private preview" in VideoFX through a waitlist, and Google said it would later bring some of Veo's capabilities to YouTube Shorts and other products.^[44] From the first day, all videos generated by Veo on VideoFX were watermarked with SynthID.^[18][44] DeepMind's Veo page described the model's "latent diffusion transformers" and said Google had added more detail to the captions of the videos in its training data.^[45]

The December 2024 competitive context

OpenAI released Sora to ChatGPT Plus and Pro subscribers in December 2024, with a version it called Sora Turbo. Fortune and TechCrunch reported that Sora could produce 1080p clips of up to 20 seconds.^[7][8][30]

Google DeepMind announced Veo 2 one week later, on December 16, 2024.^[7] The announcement was a Google blog post by Aäron van den Oord (Research Scientist, Google DeepMind, and first author of the WaveNet and VQ-VAE papers) and Elias Roman (Senior Director, Product Management, Google Labs).^[1][59][60] The same post announced an Imagen 3 update and Whisk, a Labs experiment in which a Gemini model writes captions of input images and feeds them to Imagen 3 to remix subjects, scenes and styles.^[1] Fortune's headline said Veo 2 "trounces OpenAI's Sora with 4K resolution." TechCrunch called the 4K and two-minute-plus figures "a theoretical advantage for now," because VideoFX, where Veo 2 was then exclusively available, capped videos at 720p and eight seconds.^[7][8]

When was Veo 2 released, and how did the rollout unfold?

Veo 2 rolled out in stages over about five months, from a waitlisted Labs tool to consumer apps and developer APIs. It was retired about 18 months after its announcement.

The YouTube Shorts integration put Veo 2 inside Dream Screen, which generates AI backgrounds for Shorts, and added a new option to generate standalone clips for any Short. YouTube said the features were launching "for anyone in the US, Canada, Australia, and New Zealand, with plans to expand to more of you in the future."^[9][28]

The April 2025 rollout placed Veo 2 in the Gemini app for Gemini Advanced subscribers, part of the Google One AI Premium plan ($19.99 per month), which Google renamed Google AI Pro in May 2025.^[29][41] Users selected Veo 2 from the model menu and received an eight-second, 720p clip in 16:9 landscape format as an MP4 file, which could be shared to TikTok, YouTube and other services through Gemini's share button.^[3][10] Google did not publish the size of the monthly limit. It said only that "there is a monthly limit on how many videos you can create, but we will notify you as you approach it." Google Workspace business and education plans were not supported at launch.^[3][10]

What can Veo 2 do?

Google DeepMind's Veo 2 page grouped the model's claimed improvements into "Enhanced realism and fidelity," "Advanced motion capabilities" and "Greater camera control options."^[2]

Resolution and duration

Google's December 2024 announcement said Veo 2 would deliver video "at resolutions up to 4K, and extended to minutes in length."^[1] Fortune reported, citing a Google spokesperson, that Veo 2's 4K clips were limited to eight seconds by default but could be extended to two minutes or more.^[7] None of the Google surfaces that documented their Veo 2 output offered 4K. Every one of them produced 720p video, at eight seconds or less per generation:

TechCrunch called Veo 2's resolution and length lead over Sora "a theoretical advantage for now."^[8] TechTarget likewise noted that VideoFX was still limited to 720p and eight seconds.^[22]

Camera language and creative control

Google presented the interpretation of cinematic terminology as a central feature. Its announcement said Veo 2 "understands the unique language of cinematography: ask it for a genre, specify a lens, suggest cinematic effects." It gave as examples a low-angle tracking shot, a close-up, an "18mm lens" for a wide-angle shot and "shallow depth of field" to blur the background.^[1] Google's sample prompts used phrases such as "Low-angle tracking shot, 18mm lens," "extreme close-up shot," "slow dolly shot," "wide, static shot," "Shot on 70mm film," "chiaroscuro lighting" and "Volumetric lighting."^[2] The first Veo had already been described as understanding terms such as "timelapse."^[44] The Gemini API's Veo documentation grouped the vocabulary the model understood as follows:^[35]

TechCrunch reported that, like the first Veo, Veo 2 could generate video from a text prompt or from text and a reference image.^[8] The Gemini API launched with both text-to-video and image-to-video, with the image used as the first frame. Whisk Animate brought image animation to Google One AI Premium subscribers in April 2025.^[4][35][11] On Vertex AI, Google previewed inpainting, camera presets and interpolation (generating the frames between a defined beginning and end) in April 2025. Last-frame control and video extension became generally available on June 23, 2025.^[38][6]

Physics, motion, and human anatomy

Google described Veo 2's main advance as "an improved understanding of real-world physics and the nuances of human movement and expression, which helps improve its detail and realism overall."^[1] It also said that video models often "hallucinate" unwanted details such as extra fingers or unexpected objects, and that "Veo 2 produces these less frequently, making outputs more realistic."^[1] TechCrunch saw samples in advance. It wrote that Veo 2 seemed "to have a strong grasp of refraction and tricky liquids, like maple syrup," but that it "can't quite clear the uncanny valley."^[8]

Coverage singled out several of Google's sample clips. Tom's Guide's writer had not yet tried the model himself. He pointed to a clip showing "someone accurately slicing a tomato," which he said no other video model could achieve, including Sora.^[12] He also chose the beekeeper clip, in which bees move around a farmer as he lifts a jar of honey, for its interplay between the bees and the beekeeper.^[12] Google's prompt for that clip asked for a "35mm lens on Kodak Portra 400 film."^[1] In the clip of a dog on a flamingo pool float, Tom's Guide said the model captured "the movement of the water, the physics involved in the weight of the dog and lighting."^[12]

Hands and on-screen text remained weak points. TechTarget reported the view of Constellation Research analyst Liz Miller that, although Google said Veo 2 was less likely to hallucinate, it could face the issue of overpromising and underdelivering. Veo 2 seemed to have reduced errors in hands, she said, but would probably still get things wrong.^[22] Google's later Veo 3 technical report said that Veo 3, "while much better than previous Veo models," was "still poor at generating text."^[46]

How did Veo 2 compare to Sora in benchmarks?

Google compared Veo 2 with Meta Movie Gen, Kling v1.5, MiniMax and Sora Turbo, using 1,003 prompts from MovieGenBench, "a benchmark dataset released by Meta." All comparisons were done at 720p. Veo 2's samples were 8 seconds long, the samples the page labels "VideoGen" 10 seconds and the other models' 5 seconds, and raters were shown the full videos.^[2] Google said Veo 2 "performs best on overall preference, and for its capability to follow prompts accurately."^[2] The charts on Google DeepMind's Veo 2 page gave these pairwise results:^[2]

The head-to-head against Sora Turbo was the most widely reported result. TechRadar gave the 58.8%, 14.5% and 26.7% split, and Fortune rounded it to 59% and 27%.^[19][7] Fortune also noted that Veo 2's share slipped slightly below 50% against Kling v1.5, although Veo 2 was still preferred more often than Kling.^[7][2] The Decoder reported that Veo 2 came out ahead of all four rivals on both overall quality and prompt adherence.^[13] These were Google's own evaluations, run on a prompt set released by Meta.^[2]

How does Veo 2 work? Architecture and training

Google DeepMind has not published a full technical paper for Veo 2. The closest primary sources describe the Veo models on either side of it. DeepMind's 2024 page for the original Veo said it used "latent diffusion transformers" and "high-quality, compressed representations of video (also known as latents)."^[45] Google's Veo 3 technical report says that Veo 3 "uses latent diffusion." Video and audio are encoded by separate autoencoders into compressed latents, and "a transformer-based denoising network" is trained to remove noise from them.^[46] Neither that report nor the Veo 3 model card describes Veo 2.^[46][14]

Latent diffusion approach

Generating video directly in pixel space is computationally expensive. A latent diffusion model instead works in a compressed latent space produced by a learned autoencoder. DeepMind said of the first Veo that using latents made the model more efficient.^[45] During training, the denoising network learns to remove noise from noisy latent vectors. At sampling time it is applied iteratively to Gaussian noise to produce a video latent, which is then decoded back into frames.^[46] In Veo 3 the denoiser is a transformer, a network built on self-attention. Google has not published layer-level details or parameter counts for any Veo model.^[46]

Conditioning and prompt understanding

For Veo 2 itself, the main training detail Google gave came from Eli Collins, DeepMind's vice president of product. He told TechCrunch that "Veo has been trained on high-quality video-description pairings," meaning videos paired with descriptions of what happens in them.^[8] For the first Veo, Google said it had "added more details to the captions of each video in its training data" to help the model follow prompts.^[45] In the Gemini API, Veo 2 accepted English-language prompts, an optional negative prompt and an optional first-frame image. A prompt rewriter (enhance_prompt) was enabled by default.^[35] Vertex AI's model page also listed English as the supported prompt language.^[5]

What data was Veo 2 trained on, and the YouTube controversy

Google has not disclosed the composition of Veo 2's training set. At launch, Fortune reported that Google DeepMind had "not said what data was used to train Veo 2." It added that Google had previously hinted that YouTube videos made up some of the training data for the original Veo.^[7] TechCrunch wrote that DeepMind would not say exactly where it obtained the videos to train Veo 2, that YouTube was "one possible source," and that DeepMind had previously said Google models like Veo "may" be trained on some YouTube content. TechCrunch also noted that DeepMind offered no mechanism for creators to remove works from its existing training sets.^[8]

In June 2025, CNBC reported that Google was turning to its catalog of 20 billion YouTube videos to train AI models including Gemini and the Veo 3 video and audio generator, citing a person not authorized to speak publicly. Google confirmed to CNBC that it uses YouTube videos for training but said it relies on "a subset" of them and honors specific agreements with creators and media companies.^[15] CNBC wrote that training on just 1% of the catalog would amount to 2.3 billion minutes of content, which experts said was more than 40 times the training data used by competing AI models.^[15] Creators and intellectual-property professionals told CNBC they had not been informed that their content could be used to train Google's models. PetaPixel reported that many creators were caught off guard.^[15][16] CNBC's report named Gemini and Veo 3, not Veo 2.^[15]

Marques Brownlee had raised the YouTube question in December 2024. In a YouTube Short, as reported by Benzinga, he described the difference between Veo 2 and rival models as "using a little bit of YouTube data" versus "owning YouTube and just using all of it."^[31]

How are Veo 2 videos watermarked? SynthID

Every video generated by Veo 2 carries an invisible SynthID watermark. In Google's words, Veo 2 outputs include "an invisible SynthID watermark that helps identify them as AI-generated."^[1] Google DeepMind introduced SynthID for video in May 2024 with the first Veo. The technique "embeds a watermark directly into the pixels of every video frame, making it imperceptible to the human eye, but detectable for identification." From that day, all videos generated by Veo on VideoFX were watermarked.^[18] Google says SynthID is designed to stand up to modifications such as "cropping, adding filters, changing frame rates, or lossy compression."^[47] Discussing its text watermark in the May 2024 post, Google said SynthID "isn't built to directly stop motivated adversaries like cyberattackers or hackers from causing harm," though it can make malicious use of AI-generated content harder, so it is not a complete defense against a determined adversarial attack.^[18] TechCrunch noted at Veo 2's launch that "like all watermarking tech, SynthID isn't foolproof."^[8]

Public verification tools came later. On May 20, 2025, Google announced SynthID Detector, a portal for checking whether content carries a SynthID watermark. It began rolling the portal out to early testers, with a waitlist for journalists, media professionals and researchers. Google said SynthID by then covered content from its Gemini, Imagen, Lyria and Veo models, and that over 10 billion pieces of content had already been watermarked.^[48] As of October 2026, Google's SynthID page says users can upload an image, video or audio clip to Gemini and ask whether Google AI made it, and that Gemini, Search and Chrome will check for a SynthID watermark. The page also says access to SynthID Detector is expanding to everyone.^[47]

In YouTube Shorts, YouTube said it used "SynthID watermarks and clear labels to indicate that these creations were generated with AI."^[9]

Where could you use Veo 2? Availability across Google surfaces

Veo 2 eventually reached most of Google's consumer, creator and enterprise products. Access, pricing and output varied by surface.

The Gemini API price was 30% below the Vertex AI price, so the same eight-second clip cost $2.80 through the Gemini API and $4.00 on Vertex AI.^[4][37]

How much did Veo 2 cost?

Google published Veo 2's prices in the Gemini API documentation, on the Vertex AI generative AI pricing page, and in the descriptions of its consumer subscriptions. The per-second prices were the same in every archived price list checked, from 2025 to the spring of 2026.

At $0.50 per second on Vertex AI, an eight-second clip cost $4.00, a minute of output $30 and an hour $1,800. At the Gemini API rate the same eight-second clip cost $2.80.^[37][4] Engadget reported the Vertex AI price on February 24, 2025 and noted that comparison with Sora was difficult because Sora was sold by subscription: the $200 ChatGPT Pro plan included 500 videos a month.^[27]

For consumers, Veo 2 was bundled into subscriptions rather than priced per clip, and Google did not publish the Gemini app's monthly cap.^[3] In June 2025, Google AI Pro subscribers received three Veo 3 Fast generations a day in the Gemini app. 9to5Google reported that once they reached that limit, they were using Veo 2 until it reset the next day.^[40]

Reception

Veo 2 received largely positive coverage, much of it framed around Sora Turbo's release the week before.^[7][8][19]

Marques Brownlee posted on December 16, 2024: "Google's new video generation model is called Veo 2, and if these hand-picked examples are real, they look better than anything I've gotten out of SORA..."^[17] Benzinga reported that in a YouTube Short a few days later, he said Sora often revealed its AI nature through inaccuracies in movement and physics, while Veo 2 delivered more realistic results. He also acknowledged flaws such as trouble with complex scenes and garbled text.^[31] Access was limited at first. Axios noted that Veo 2 was available only to early-access users in the U.S. who were over 18, and quoted the tech writer MG Siegler: "No one seems to have any idea how to actually try it out."^[30]

Fortune's headline said Veo 2 "trounces OpenAI's Sora with 4K resolution."^[7] The Decoder reported the MovieGenBench results under the headline that Veo 2 "outperforms OpenAI's Sora Turbo."^[13] TechRadar said Veo 2 "beats OpenAI Sora with 4K AI video generation," and SiliconANGLE led with Veo 2's ability to generate high-resolution videos up to two minutes long.^[19][20] Tom's Guide called Veo 2 "one of the best AI video models I've ever seen," although its writer had not yet tried the model.^[21] In January 2025 Axios wrote that "there's no agreed-upon benchmarking system for AI video generators," but that Veo 2 was "widely seen as the most promising right now."^[30]

More measured assessments came from outlets that focused on enterprise use. TechTarget's coverage was titled "Google's Veo 2 is technically advanced, but concerns remain." It reported Liz Miller's view that Google had not been clear about Veo 2's training data, although it had likely used YouTube. It also reported that Google had not included Veo 2 users in its contractual indemnification clauses, and that VideoFX was still limited to 720p and eight seconds.^[22] TechCrunch reported that Google's indemnity policy would not apply to Veo 2 until the model was generally available.^[8]

What was Veo 2 used for?

Google and its partners described several commercial uses of Veo 2 in 2025.

On April 2, 2025, Envato announced VideoGen, a feature built with Veo 2 on Vertex AI that turns text or images into video. Envato said tens of thousands of subscribers had used it in its first few days. Paid subscribers received 10 generations a month of videos up to five seconds long.^[23]

In an April 2025 Google Cloud blog post, Google cited two customers. Goodby Silverstein & Partners and the Dalí Museum used Veo 2 and Imagen 3 to realize "Giraffes on Horseback Salad," a film Salvador Dalí imagined in 1937 that was never produced. Kraft Heinz's Tastemaker platform gave its teams access to Veo 2 and Imagen 3.^[38] In its Gemini API launch post, Google said the game company Wolf Games was using Veo 2 in a generative platform for personalized interactive story games.^[4]

Google Vids added Veo 2 on May 6, 2025. Workspace users in eligible editions could generate eight-second, 720p clips directly in Vids "without extra costs," subject to daily limits.^[24]

What are Veo 2's limitations?

Google's own Veo 2 page acknowledged that "creating realistic, dynamic, or intricate videos, and maintaining complete consistency throughout complex scenes or those with complex motion, remains a challenge."^[2] Eli Collins told TechCrunch that "Coherence and consistency are areas for growth." He added that "character consistency can be a challenge" and that there was "room to improve in generating intricate details, fast and complex motions."^[8]

Clip length was the most visible practical limit. Google described Veo 2 as capable of minutes-long video, but the Google products and APIs that documented their output produced at most eight seconds per generation.^{[1][3][35][5]} Vertex AI later added video extension for Veo 2, generally available from June 23, 2025. Veo 3.1 brought scene extension to the Gemini API and Flow in October 2025.^[6][26]

Hands and text rendering remained weak points, as described above.^[22][46]

Veo 2 generated video only, without sound. Google described Veo 3 as "our first video model to incorporate high-fidelity video outputs and native audio," so any Veo 2 clip that needed speech or sound required audio from another tool.^[42]

Access varied by region. VideoFX initially rolled out only in the U.S., YouTube's Veo 2 features launched in four countries, and Whisk Animate launched in more than 60 countries.^[7][9][11] The Gemini API documentation also listed the limitation "Image-to-video personGeneration is not allowed in EU, UK, CH, MENA locations."^[35]

Content controls were built into the API. A personGeneration setting in the Gemini API controlled whether videos could include people: "dont_allow" did not allow people or faces, and "allow_adult" generated adults but not children.^[35] TechCrunch reported that DeepMind used prompt-level filters, including for violent, graphic and explicit content.^[8]

How does Veo 2 differ from Veo 1?

The table below compares the two models using Google's May 2024 and December 2024 announcements and later documentation.

Google also said Veo 2 improved on physics and human movement and produced fewer hallucinated details such as extra fingers.^[1]

Successors: Veo 3 and Veo 3.1

Veo 3 (May 2025)

Veo 3 was announced on May 20, 2025 at Google I/O 2025 as the next major iteration after Veo 2.^[43] Its defining new capability was native audio. Google said Veo 3 natively generates dialogue, effects and music and "synchronizes it with video in a single pass." It described Veo 3 as "our first video model to incorporate high-fidelity video outputs and native audio."^[42] At a press briefing, Demis Hassabis said, "For the first time, we're emerging from the silent era of video generation." TechCrunch called audio a potential "big differentiator" for Veo 3, while noting that "AI-powered sound-generating tools aren't novel, nor are models to create video sound effects."^[43]

Veo 2 was not withdrawn when Veo 3 launched. Google's May 20, 2025 release notes said Google AI Pro subscribers would "retain access" to video generation with Veo 2 in the Gemini app, while Veo 3 went to the new Google AI Ultra plan. Pro subscribers also got Flow, where Google said they could "try the tool with our Veo 2 model."^[41][39] On June 9, 2025, Pro subscribers received three Veo 3 Fast generations a day in the Gemini app and used Veo 2 after reaching that limit. At that point Flow offered Veo 2 in "Fast" and "Quality" modes, at 10 and 100 credits per generation, alongside Veo 3 Fast at 20 credits.^[40]

Veo 3.1 (October 2025)

Veo 3.1 was launched on October 15, 2025 alongside updates to the Flow filmmaking tool. It was released in paid preview in the Gemini API and also came to Vertex AI and the Gemini app.^[25][26] Google said it "builds on Veo 3."^[25] The additions most relevant to a comparison with Veo 2 included reference-image conditioning ("Ingredients to Video," with up to three reference images of a character, object or scene) and first-and-last-frame transitions "complete with accompanying audio."^[26] They also included scene extension that could create videos "lasting for a minute or more," with each new clip generated from the final second of the previous one, and richer native audio.^[25][26]

Google said "Veo 3.1 is the same price as Veo 3."^[26] By November 2025 the Gemini API listed Veo 3.1 and Veo 3 at $0.40 per second and their Fast variants at $0.15 per second, all with audio, against $0.35 for the silent Veo 2.^[36] Veo 2 was therefore no longer the cheapest Veo option in the Gemini API. On Vertex AI, its $0.50 rate was above the October 2025 prices for Veo 3 ($0.40 with audio, $0.20 without) and Veo 3 Fast ($0.15 with audio, $0.10 without).^[37]

Retirement

Veo 2 had no announced end date for most of its life. As late as February 2026, Google's Vertex AI lifecycle table listed "No retirement date announced" for veo-2.0-generate-001.^[32] On March 24, 2026, the Vertex AI release notes listed it among deprecated video generation endpoints. They recommended moving to veo-3.1-generate-001 before June 30, 2026 "to avoid service disruption."^[6] In the Gemini API, a June 15, 2026 changelog entry announced that veo-2.0-generate-001 would be shut down on June 30, 2026, along with the stable Veo 3 and Veo 3 Fast models. It told developers to move to the Veo 3.1 preview models or to the Veo 3.1 GA models on the Gemini Enterprise Agent Platform.^[34]

Both retirements took effect on June 30, 2026. The Gemini API deprecations page lists "gemini-omni-1.1-flash or the GA models on the Gemini Enterprise Agent Platform" as the replacement. The Agent Platform's lifecycle table lists Veo 2 among retired models, with veo-3.1-generate-001 as the recommended upgrade.^[33][32] Google had announced Gemini Omni on May 19, 2026, and its Gemini app video page says "Gemini Omni will replace Veo in the Gemini app."^[57][58]

Comparison with peer models

The table below compares Veo 2 with the models in Google's benchmark and with other widely covered video generators of the period. It uses what developers and contemporary reporting said at the time, and it omits MiniMax, which Google rated but whose launch details are not covered by the sources cited here. Output limits changed often after launch.

Veo 2 never generated sound. Native audio spread across the field during 2025:

At launch, the most direct competitor was Sora Turbo. Fortune and TechCrunch contrasted Veo 2's claimed 4K resolution with Sora's 1080p limit, while Sora's 20-second clips were longer than the eight seconds VideoFX allowed.^[7][8] Fortune also contrasted the two marking schemes. Veo 2 used invisible SynthID watermarks, while Sora embedded "a visible animation in the bottom right corner" of its videos and used the C2PA protocol.^[7] Runway released Gen-4 in March 2025, emphasizing consistent characters and scenes from reference images.^[54] In March 2026 OpenAI announced that it would discontinue Sora, closing the web and app versions on April 26, 2026 and the Sora API on September 24, 2026.^[56]

Meta's Movie Gen, announced on October 4, 2024, was the source of MovieGenBench, the prompt set Google used for its comparison.^[55][2] Meta presented Movie Gen as "generative AI research for media" and said its models were trained "on a combination of licensed and publicly available datasets."^[55]

Legacy and current status

Within about five months of its announcement, Veo 2 was available in YouTube Shorts, the Gemini app, Whisk, Google Vids, the Gemini API and Vertex AI.^{[9][3][24][4][5]} Its December 2024 result against Sora Turbo (58.8% to 26.7% in Google's own evaluation) was widely reported, by TechRadar in full and by Fortune in rounded form.^[19][7] The gap between Google's "up to 4K" and "minutes in length" description and the 720p, eight-second clips that its products delivered was noted from the first day.^[8][22]

Veo 2 carried the SynthID watermarking that Google had applied to Veo videos since May 2024.^[18] Google's Gemini app page says Gemini Omni "will replace Veo in the Gemini app," and Google says all videos created with Omni include the SynthID watermark.^[58][57]

Veo 2 was retired on Vertex AI and shut down in the Gemini API on June 30, 2026.^[32][33] As of October 2026, Google's Agent Platform lifecycle table still lists veo-3.1-generate-001 and veo-3.1-fast-generate-001, with retirement dates of "November 17, 2026 or later." In the Gemini API, the Veo 3.1 preview models were shut down on October 22, 2026, with gemini-omni-1.1-flash listed as their replacement.^[32][33]

Continue reading

• VeoVeo is a family of text-to-video generative AI models developed by Google DeepMind. The original Veo was announced at Google I/O on May 14, 2024, and TechCrunch described it as Google's answer to OpenAI's Sora.Mentioned 234 times above
• GoogleGoogle LLC is a United States technology company and the largest operating business within Alphabet Inc. Its activities include internet search, online advertising, video, mobile and web platforms, consumer…Mentioned 141 times above
• Gemini (language model)Gemini is a family of natively multimodal large language models developed by Google DeepMind, first announced on December 6, 2023, that can reason across text, images, audio, video, and code within a single…Mentioned 76 times above
• Gemini (app)The Gemini app is Google's consumer AI assistant, a chat application available on the web at gemini.google.com and as native Android, iOS, macOS and Windows appsMentioned 76 times above
• Veo 3Veo 3 is a video generation model developed by Google DeepMind and announced at Google I/O on May 20, 2025.Mentioned 47 times above
• GGoogle Vertex AIGoogle Vertex AI is the unified machine learning and generative artificial intelligence platform offered by Google Cloud, announced at Google I/O on May 18Mentioned 38 times above

More from See also

• Veo 3.1
• Gemini Omni
• Sora
• Sora 2
• Runway (company)
• Runway Gen-3 Alpha
• Runway Gen-4
• Kling (video generation)
• Imagen 3
• Imagen 4
• Whisk (Google Labs)
• Google DeepMind
• SynthID
• Stable Video Diffusion
• Pika (video generation)
• Demis Hassabis
• Diffusion model
• Text-to-video generation
• Generative AI

Topics:Generative AIGoogle DeepMindVideo Generation

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24Aäron van den Oord and Elias Roman. "State-of-the-art video and image generation with Veo 2 and Imagen 3." Google Blog, December 16, 2024. blog.google/...age-generation-update-december-2024
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Google DeepMind. "Veo 2." Technologies page with benchmark charts, archived December 16, 2024. web.archive.org/...veo-2
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Google Blog. "Try generating video in Gemini, powered by Veo 2." April 15, 2025. blog.google/...video-generation
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Google Developers Blog. "Gemini 2.5 Flash and Pro, Live API, and Veo 2 in the Gemini API." April 9, 2025. developers.googleblog.com/...-api-veo-2-gemini-api
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Google Cloud Documentation. "Veo on Vertex AI API" reference, archived February 7, 2025 web.archive.org/...veo-video-generation and April 11, 2025 web.archive.org/...veo-video-generation ; "Veo 2 Generate 001" model page, archived July 19, 2025 web.archive.org/...2-0-generate-001
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Google Cloud Documentation. "Generative AI on Vertex AI release notes" (entries of June 23, 2025, October 15, 2025 and March 24, 2026; accessed October 30, 2026). docs.cloud.google.com/...release-notes
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19David Meyer. "Google DeepMind's new Veo 2 AI video generator trounces OpenAI's Sora with 4K resolution." Fortune, December 16, 2024. fortune.com/...2-ai-video-generator-4k-openai-sora
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24Kyle Wiggers. "Google DeepMind unveils a new video model to rival Sora." TechCrunch, December 16, 2024. techcrunch.com/...-a-new-video-model-to-rival-sora
9. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9YouTube Blog. "Imagine it, create it: Veo 2 is coming to YouTube Shorts." February 13, 2025. blog.youtube/...veo-2-shorts
10. ^1 ^2Kyle Wiggers. "Google's Veo 2 video generating model comes to Gemini." TechCrunch, April 15, 2025. techcrunch.com/...-video-generator-comes-to-gemini
11. ^1 ^2 ^3 ^4 ^5Google Blog. "Generate videos in Gemini and Whisk with Veo 2." April 15, 2025. blog.google/...video-generation
12. ^1 ^2 ^3Ryan Morrison. "Google Veo 2 is one of the best AI video models I've ever seen, here's 5 examples of what it can do" (headline uses a dash). Tom's Guide, December 17, 2024. tomsguide.com/...eres-5-examples-of-what-it-can-do
13. ^1 ^2Matthias Bastian. "Google's Veo 2 outperforms OpenAI's Sora Turbo in head-to-head AI video generation tests." The Decoder, December 16, 2024. the-decoder.com/...-head-ai-video-generation-tests
14. ^Google. "Veo 3 model card." 2025. storage.googleapis.com/...Veo-3-Model-Card.pdf
15. ^1 ^2 ^3 ^4Zach Vallese. "Google is using YouTube videos to train its Gemini, Veo 3 AI models." CNBC, June 19, 2025. cnbc.com/...google-youtube-ai-training-veo-3
16. ^Matt Growcoot. "YouTubers Surprised That Google Uses Their Videos to Train AI Models." PetaPixel, June 20, 2025. petapixel.com/...deos-to-train-its-ai-models-veo-3
17. ^Marques Brownlee on X. December 16, 2024. x.com/...1868764927856722295
18. ^1 ^2 ^3 ^4 ^5 ^6Google DeepMind. "Watermarking AI-generated text and video with SynthID." May 14, 2024. deepmind.google/...ted-text-and-video-with-synthid
19. ^1 ^2 ^3 ^4 ^5John-Anthony Disotto. "Google's new Veo 2 beats OpenAI Sora with 4K AI video generation, here's how to try it" (headline uses a dash). TechRadar, December 17, 2024. techradar.com/...eo-generation-heres-how-to-try-it
20. ^Maria Deutscher. "Google debuts Veo 2 video generator, upgraded Imagen 3 with Whisk remix tool." SiliconANGLE, December 16, 2024. siliconangle.com/...eo-generator-upgraded-imagen-3
21. ^Ryan Morrison. "Google Veo 2 is one of the best AI video models I've ever seen, here's 5 examples of what it can do" (headline uses a dash). Tom's Guide, December 17, 2024. tomsguide.com/...eres-5-examples-of-what-it-can-do
22. ^1 ^2 ^3 ^4 ^5Esther Shittu. "Google's Veo 2 is technically advanced, but concerns remain." TechTarget, December 18, 2024. techtarget.com/...lly-advanced-but-concerns-remain
23. ^1 ^2Google Cloud Press Corner. "Envato Launches Video Generation Feature for Creatives, Developed with Veo 2 on Vertex AI." April 2, 2025. googlecloudpresscorner.com/...h-Veo-2-on-Vertex-AI
24. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Google Workspace Updates. "Generate custom video clips in Google Vids with Veo 2." May 6, 2025. workspaceupdates.googleblog.com/...-in-google-vids
25. ^1 ^2 ^3 ^4Google Blog. "Introducing Veo 3.1 and advanced capabilities in Flow." October 15, 2025. blog.google/...veo-updates-flow
26. ^1 ^2 ^3 ^4 ^5Google Developers Blog. "Introducing Veo 3.1 and new creative capabilities in the Gemini API." October 15, 2025. developers.googleblog.com/...ies-in-the-gemini-api
27. ^1 ^2 ^3Jeremy Gan. "Each second of Google's Veo 2 AI video footage will cost 50 cents." Engadget, February 24, 2025. engadget.com/...otage-will-cost-50-cents-155419146
28. ^1 ^2Amanda Silberling. "YouTube Shorts adds Veo 2 so creators can make GenAI videos." TechCrunch, February 13, 2025. techcrunch.com/...-creators-can-make-gen-ai-videos
29. ^1 ^2 ^3 ^4 ^5Abner Li. "Gemini app rolling out Veo 2 video generation for Advanced users." 9to5Google, April 15, 2025. 9to5google.com/...gemini-app-veo-2
30. ^1 ^2 ^3 ^4 ^5 ^6Axios. "AI video tools: OpenAI's Sora, Google's Veo 2 and Runway's Gen-3 Alpha compared." January 3, 2025. axios.com/...openai-sora-google-veo-runway-video ; archived web.archive.org/...ai-sora-google-veo-runway-video
31. ^1 ^2Ananya Gairola. "Marques Brownlee Breaks Down How Google's AI Video Generator Veo2 Outpaces OpenAI's Sora." Benzinga, December 20, 2024. benzinga.com/...nerator-veo2-outpaces-openais-sora
32. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Google Cloud. "Model versions and lifecycle." Gemini Enterprise Agent Platform documentation (accessed October 30, 2026). docs.cloud.google.com/...model-versions ; Vertex AI version archived February 8, 2026 web.archive.org/...model-versions and May 14, 2026 web.archive.org/...model-versions
33. ^1 ^2 ^3 ^4 ^5Google AI for Developers. "Gemini deprecations" (accessed October 30, 2026). ai.google.dev/...deprecations
34. ^1 ^2 ^3 ^4Google AI for Developers. "Release notes" (Gemini API changelog; entries of April 9, 2025, October 15, 2025 and June 15, 2026; accessed October 30, 2026). ai.google.dev/...changelog
35. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18Google AI for Developers. "Generate video using Veo." Gemini API documentation, archived May 16, 2025. web.archive.org/...video
36. ^1 ^2 ^3 ^4Google AI for Developers. "Gemini Developer API pricing." Archived May 11, 2025 web.archive.org/...pricing ; November 3, 2025 web.archive.org/...pricing ; May 1, 2026 web.archive.org/...pricing
37. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Google Cloud. "Vertex AI pricing" (generative AI). Archived March 2, 2025 web.archive.org/...pricing ; August 6, 2025 web.archive.org/...pricing ; October 2, 2025 web.archive.org/...pricing ; April 1, 2026 web.archive.org/...pricing
38. ^1 ^2 ^3Warren Barkley. "Expanding generative media for enterprise on Vertex AI." Google Cloud Blog, April 2025. cloud.google.com/...ia-for-enterprise-on-vertex-ai
39. ^1 ^2 ^3 ^4 ^5 ^6 ^7Shimrit Ben-Yair. "Introducing Google AI Ultra: The best of Google AI in one subscription." Google Blog, May 20, 2025. blog.google/...google-ai-ultra
40. ^1 ^2 ^3 ^4 ^5 ^6Abner Li. "Gemini app Pro subscribers now get three 'Veo 3 Fast' videos per day." 9to5Google, June 9, 2025. 9to5google.com/...gemini-veo-3-fast
41. ^1 ^2 ^3 ^4Google. "Gemini Apps' release updates & improvements" (entries of 2025.04.22 and 2025.05.20). gemini.google/release-notes
42. ^1 ^2 ^3 ^4 ^5 ^6 ^7Google Developers Blog. "Build with Veo 3, now available in the Gemini API." July 17, 2025. developers.googleblog.com/...-available-gemini-api
43. ^1 ^2 ^3Kyle Wiggers. "Veo 3 can generate videos, and soundtracks to go along with them" (headline uses a dash). TechCrunch, May 20, 2025. techcrunch.com/...oundtracks-to-go-along-with-them
44. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Eli Collins and Douglas Eck. "New generative media models and tools, built with and for creators." Google Blog, May 14, 2024. blog.google/...google-generative-ai-veo-imagen-3
45. ^1 ^2 ^3 ^4Google DeepMind. "Veo." Technologies page, archived May 31, 2024. web.archive.org/...veo
46. ^1 ^2 ^3 ^4 ^5 ^6Google DeepMind. "Veo 3 Technical Report." 2025. storage.googleapis.com/...Veo-3-Tech-Report.pdf
47. ^1 ^2Google DeepMind. "SynthID" (accessed October 30, 2026). deepmind.google/...synthid
48. ^Pushmeet Kohli. "SynthID Detector, a new portal to help identify AI-generated content" (headline uses a dash). Google Blog, May 20, 2025. blog.google/...google-synthid-ai-content-detector
49. ^Amanda Silberling. "OpenAI is launching the Sora app, its own TikTok competitor, alongside the Sora 2 model." TechCrunch, September 30, 2025. techcrunch.com/...titor-alongside-the-sora-2-model
50. ^Kuaishou Technology. "Kling AI Launches Video 2.6 Model with 'Simultaneous Audio-Visual Generation' Capability, Redefining AI Video Creation Workflow." PR Newswire via Nasdaq, December 5, 2025. nasdaq.com/...s-audio-visual-generation-capability
51. ^Ivan Mehta and Rebecca Bellan. "Runway releases its first world model, adds native audio to latest video model." TechCrunch, December 11, 2025. techcrunch.com/...tive-audio-to-latest-video-model
52. ^1 ^2Maximilian Schreiner. "AI video generator KLING releases version 1.5 with impressive features." The Decoder, September 19, 2024. the-decoder.com/...on-1-5-with-impressive-features
53. ^1 ^2Anastasis Germanidis. "Runway Gen-3 Alpha: AI Video Generation with Temporal Control." Runway, June 17, 2024. runway.com/...introducing-gen-3-alpha
54. ^1 ^2 ^3Kyle Wiggers. "Runway releases an impressive new video-generating AI model." TechCrunch, March 31, 2025. techcrunch.com/...ve-new-video-generating-ai-model
55. ^1 ^2 ^3 ^4Meta AI. "How Meta Movie Gen could usher in a new AI-enabled era for content creators." October 4, 2024. ai.meta.com/...undation-models-generative-ai-video
56. ^Matthias Bastian. "OpenAI sets two-stage Sora shutdown with app closing April 2026 and API following in September." The Decoder, March 28, 2026. the-decoder.com/...-and-api-following-in-september
57. ^1 ^2Koray Kavukcuoglu. "Introducing Gemini Omni." Google Blog, May 19, 2026. blog.google/...gemini-omni
58. ^1 ^2Google. Gemini app video generation page (accessed October 30, 2026). gemini.google/...video-generation
59. ^Aaron van den Oord et al. "WaveNet: A Generative Model for Raw Audio." arXiv:1609.03499, September 2016. arxiv.org/...1609.03499
60. ^Aaron van den Oord, Oriol Vinyals and Koray Kavukcuoglu. "Neural Discrete Representation Learning" (VQ-VAE). arXiv:1711.00937, November 2017. arxiv.org/...1711.00937
61. ^1 ^2 ^3Google Developers Blog. "Veo 3 and Veo 3 Fast, new pricing, new configurations and better resolution" (headline uses a dash). September 8, 2025. developers.googleblog.com/...and-better-resolution

Show all 61 references