Veo 3

A video model that also makes sound

Veo 3 is a video generation model that Google announced at Google I/O on 20 May 2025 ^[1]. The capability Google put at the centre of the announcement is sound: the model generates video with native audio, including things like traffic noise, birdsong and dialogue between characters, with the speech matched to the movement of their lips ^[1]. Google described Veo 3 as able, for the first time, to also generate videos with audio ^[1].

That combination matters because of how much the output has to hold together. A silent clip only has to look plausible frame by frame; a clip with dialogue has to keep two things consistent at once — the word being spoken and the shape of the mouth saying it. When audio and video come out of one model, the timing of both has to be decided together, which is why lip syncing is the part of the announcement easiest to check by eye.

Where it was available

At launch, Veo 3 was offered to Google AI Ultra subscribers in the United States through the Gemini app and through Flow, and to enterprises through Vertex AI ^[1]. Flow is an AI filmmaking tool built on Veo, Imagen and Gemini, which Google launched for Google AI Pro and Ultra subscribers in the United States ^[1]. Within Google's own products, Veo 3 was used in Flow and in the Gemini app ^[2].

The split is typical of how generation models reach users: a consumer app where one prompt produces a clip, a filmmaking tool where clips are assembled into longer sequences, and a cloud platform where developers call the model directly.

Watermarks

Veo 3 outputs carry SynthID watermarks, and Google launched a SynthID Detector portal on the same day ^[1]. SynthID is Google's scheme for embedding a signal in generated media that can later be detected; the detector portal is the matching tool for checking whether a file carries it. For a model whose output can contain speech and realistic scenes, marking the result is one way to keep generated clips distinguishable from recorded footage.

Timeline in the API

The model's path through the Gemini API is recorded in Google's deprecation table. A preview model, veo-3.0-generate-preview, appeared on 31 July 2025; the generally available veo-3.0-generate-001 and veo-3.0-fast-generate-001 followed on 9 September 2025, and both were shut down on 30 June 2026 ^[3]. The preview models were retired earlier: veo-3.0-generate-preview and veo-3.0-fast-generate-preview were shut down on 12 November 2025 ^[3]. For the two generally available models, the deprecation table lists gemini-omni-1.2-flash or a generally available model on Gemini Enterprise Agent Platform as replacements ^[3].

The sequence — a preview, then a stable version alongside a faster variant, then a shutdown date with a named successor — is how API models are commonly managed. A preview lets developers try the model while its interface can still change; the stable name is the one applications are expected to build on; the shutdown date and the replacement point to where the provider wants callers to move.

Where it sits

Veo 3 belongs to the line of video generation models that produced silent clips; what it adds is an audio track generated together with the picture ^[1]. The launch also shows the surrounding pieces a model of this kind now comes with: a consumer app, a filmmaking tool, an enterprise platform, a watermark and a detector for it, and a published schedule for API versions. What a user actually meets is often those pieces rather than the model alone.