Veo 2 is the second generation of Google DeepMind’s video model, announced in December 2024. It was the first Veo most people could actually touch: where the original Veo stayed a research preview, Veo 2 rolled out through VideoFX in Google Labs, then into the Gemini app, Whisk, and the Vertex AI API. For about five months it was Google’s answer to Sora — praised for physics and camera language, and famously silent, because it generated no audio.
That era is over. Veo 3 arrived in May 2025 with native sound, Veo 3.1 followed in October 2025, and Veo 4 is now the flagship. Google has moved its apps and API onto the newer generations, which is why so many "veo 2" searches now land on dead or redirected pages.
This guide keeps the record straight: what Veo 2 could really do, where it ran, how it differs from Veo 3, and the practical upgrade path — including why Veo AI carries Veo 3.1 and Veo 4 rather than Veo 2, and what that means if you arrived here holding an old bookmark.
What was Veo 2?
Veo 2 was Google DeepMind’s text-to-video and image-to-video model, unveiled on December 16, 2024 as the successor to the original Veo shown at I/O that spring. Its pitch was realism: markedly better physics than the first generation, a genuine understanding of cinematography terms — lenses, dolly moves, depth of field — and fewer of the melting hands and morphing objects that defined early AI video.
It also marked a strategy shift. The first Veo was a demo; Veo 2 was a product. Google put it in front of consumers through VideoFX at labs.google, let Gemini subscribers generate clips in the app, wired it into the Whisk experiment as "Whisk Animate" to bring still images to life, and sold it to developers through the API. If someone used "Google AI video" in early 2025, they used Veo 2.
What Veo 2 could actually do
The honest capability sheet is more modest than the launch coverage suggested. DeepMind’s research announcement promised resolution up to 4K and clips extending past two minutes; what shipped to users was mostly 8-second clips at 720p, with 1080p arriving in limited surfaces later. The gap between the research ceiling and the product reality became a running theme in Veo 2 reviews.
- Text to video: 8-second clips from a prompt, typically at 720p in the shipped products.
- Image to video: animate a still — the feature Whisk Animate was built on, and a big reason "veo 2 image to video" was so searched.
- Camera control via language: prompts specifying shot types, lens behavior, and movement were followed unusually well for the time.
- Improved physics: water, fabric, and motion held together far better than first-generation AI video.
- No native audio: every Veo 2 clip was silent; soundtracks had to be added in an editor afterwards.
- SynthID watermarking: all output carried Google’s invisible provenance mark.
Where you could use Veo 2
Access ran through four doors, and all four have since moved on. VideoFX in Google Labs was the original waitlisted home. The Gemini app offered Veo 2 generation to paying subscribers from early 2025. Whisk animated still images with it. And Vertex AI plus the Gemini API carried it for developers — the "veo2 api" that older integration tutorials reference.
Today each of those surfaces runs a newer generation: Flow superseded VideoFX as Google’s filmmaking tool, Gemini generates with current Veo models, and the API lists Veo 2 only as a legacy entry where it appears at all. Searching "try veo 2" in 2026 is effectively searching for a museum ticket.
Veo 2 vs Veo 3: what actually changed
The jump from Veo 2 to Veo 3 in May 2025 was the largest single-generation leap Veo has made, and one feature defines it: audio. Veo 3 generates dialogue, ambient sound, and effects natively, in sync with the picture — turning silent animations into finished shots. Lip-synced speech from a text prompt was the demo that made Veo 3 a phenomenon, and it was structurally impossible on Veo 2.
Beyond sound, Veo 3 improved prompt adherence and temporal consistency, and Veo 3.1 then layered on the control surface Veo 2 users had been asking for: reference images for consistent characters, first-and-last-frame keyframes, clip extension, and richer audio. Veo 4 pushed fidelity and realism further still. In every dimension Veo 2 competed on, its successors are simply better — there is no niche where Veo 2 remains the right pick.
The Veo timeline at a glance
Five generations in under two years. Here is the family history that search results tend to scramble.
| Generation | Released | Headline change | Status in 2026 |
|---|---|---|---|
| Veo 1 | May 2024 | Research debut at Google I/O | Never widely released |
| Veo 2 | December 2024 | First broad availability — VideoFX, Gemini, API; no audio | Retired from Google’s main surfaces |
| Veo 3 | May 2025 | Native audio: dialogue, ambience, effects | Superseded by 3.1 |
| Veo 3.1 | October 2025 | Reference images, keyframes, extension, richer audio | Current workhorse — on Veo AI |
| Veo 4 | 2026 | Flagship fidelity and realism | Current flagship — on Veo AI |
Can you still use Veo 2 — and should you?
Practically, no — and honestly, you should not want to. Google’s consumer products no longer offer Veo 2 generation, and Veo AI does not carry it either: our catalog runs Veo 4 Pro and Fast, Veo 3.1 Pro and Fast, and Veo 3.1 Lite, alongside Sora 2, Kling, Hailuo, Wan, Seedance, and 40+ other models. Listing a silent, superseded model would cost you credits for strictly worse output, so we do not.
The upgrade path is painless. Veo 3.1 is the closest current equivalent to what Veo 2 did — text to video and image to video — and improves on it everywhere: native audio, 1080p available, 4/6/8-second durations, 16:9 and 9:16, reference images, keyframes, negative prompts, and extension. Prompts written for Veo 2 run on Veo 3.1 unchanged; they simply come back with sound. On Veo AI you also get watermark-free downloads, free starter credits on sign-up, and the credit cost of each clip shown before you generate.