# Text-to-Video > How text-to-video models work, what they can do in late 2026 (audio, 30-second clips, references), what they cost, and the year Sora and Veo were replaced. Source: https://metavert.io/text-to-video Published: 2026-10-21 Updated: 2026-10-21 **Text-to-video** is the generation of a video clip from a written prompt. Its sibling, image-to-video, starts from a supplied frame or a set of reference images. Both are the core capabilities of [generative video](https://metavert.io/generative-video). By late 2026 the leading models generate synchronized sound as well as picture, produce single-pass clips of up to 30 seconds, hold characters consistent across multiple shots, accept dozens of reference images and clips, and can revise a clip in conversation. They have also become one of the fastest-churning markets in AI. Two of the best-known names of 2024 and 2025, Sora and Veo, have been discontinued or superseded within 2026. ### How It Works Most current models are latent [diffusion models](https://metavert.io/diffusion-models) built on transformers. A video autoencoder compresses footage into a much smaller [latent space](https://metavert.io/latent-space) that spans both space and time. The latent is cut into patches that act as tokens, and a [diffusion transformer](https://metavert.io/diffusion-transformers) (DiT) learns to turn noise into a clean latent, guided by an embedding of the text prompt. The decoder then turns the latent back into frames. The DiT design comes from Peebles and Xie's 2022 paper "Scalable Diffusion Models with Transformers." OpenAI's February 2024 Sora report popularized the idea of "spacetime patches." Alibaba's open Wan report describes the same combination of a diffusion transformer with a new video autoencoder. Three refinements define the 2026 generation: - **Joint audio.** Instead of adding sound afterward, models generate it with the picture. Lightricks' LTX-2 paper, for example, describes two linked [transformer](https://metavert.io/transformer-architecture) streams, 14 billion parameters for video and 5 billion for audio, connected by cross-attention in both directions. - **Autoregressive extension.** [Autoregressive](https://metavert.io/autoregressive-generation) or hybrid approaches generate video in chunks that condition on what came before, which is how clips are extended past a model's single-pass length. - **Composite pipelines.** Some products chain models. xAI's documentation says text-to-video on Grok Imagine 1.5 "runs as text-to-image then image-to-video under the hood," which links this field closely to [text-to-image](https://metavert.io/text-to-image). ### What the Models Do in Late 2026 - **Length.** Alibaba's Wan 3.0 and ByteDance's Seedance 2.5 both make up to 30 seconds in one pass, and Seedance offers a long-video mode of up to three minutes. Kling 4.0, launched in mid-October 2026, also generates up to 30 seconds in a single pass, according to The Decoder. - **Sound.** Google's Gemini Omni Flash generates synchronized audio by default. Kling 4.0 adds stereo audio and lip-sync in more than 10 languages, and Black Forest Labs' FLUX 3 offers lip-sync in more than 14. - **References.** Seedance 2.5 accepts up to 50 references (30 images, 10 video clips and 10 audio clips) to fix characters, settings and voices. - **Conversational editing.** Gemini Omni works in multiple turns, so a clip can be changed with follow-up instructions rather than regenerated from scratch. - **Resolution.** 720p and 1080p are standard. Some 4K output is upscaled; Google notes this for Omni's 1080p and 4K modes. ### The Market's Churn 2026 rearranged the field: - **[Sora](https://metavert.io/sora) is gone.** OpenAI announced the shutdown on March 24, 2026, closed the app on April 26 and removed the Videos API and every Sora 2 model on September 24. - **Google moved from Veo to Gemini Omni.** Gemini Omni Flash was announced at Google I/O in May 2026, and Gemini Omni 1.1 Flash became generally available on August 27. Google has shut down its Veo 2 and Veo 3.0 API models, and the Veo 3.1 previews on the Gemini API end on October 22, 2026, with developers pointed to Omni. - **Chinese models lead the leaderboards.** On Artificial Analysis's blind-vote text-to-video board (with audio, as of October 20, 2026), the top five are Wan 3.0, Utopai X, Seedance 2.5, FLUX 3 Max and MiniMax H3, within 15 Elo points of each other. Three of the five come from Alibaba, ByteDance and MiniMax. Gemini Omni 1.1 Flash ranks ninth and Veo 3.1 twentieth. Kling 4.0 was too new to be listed. MiniMax H3 leads the separate image-to-video board. - **Incumbents diversified.** [Runway](https://metavert.io/runway)'s Gen-4.5 gained native audio and one-minute multi-shot output in December 2025, and the company added the Aleph 2.0 editing model in May 2026. [Kling](https://metavert.io/kling) released version 4.0. Head-to-head comparisons are in [Gemini Omni vs Runway](https://metavert.io/compare/gemini-omni-vs-runway), [Seedance vs Kling](https://metavert.io/compare/seedance-vs-kling) and [Wan vs Seedance](https://metavert.io/compare/wan-vs-seedance). ### What It Costs Prices are quoted per second of output, and they span more than an order of magnitude. On Google's Gemini API, Veo 3.1 costs $0.40 per second at 720p or 1080p and $0.60 at 4K, and Veo 3.1 Lite costs $0.05 to $0.08. Gemini Omni 1.1 Flash bills video as output tokens, a first for a major provider: 5,792 tokens per second of 720p video at $17.50 per million tokens, or about $0.10 per second. Alibaba lists Wan 3.0 at $0.05 to $0.20 per second. Because most usable shots take several attempts, the effective cost per kept second is a multiple of the list price. ### Open Weights Fewer frontier video models ship [open weights](https://metavert.io/open-weight-models) than language models, and the licenses vary widely: - **Lightricks LTX-2.5** (22 billion parameters, native audio) allows commercial use under its community license for companies under $10 million in annual revenue. - **Wan 2.2** (Apache-2.0) is still the newest open Wan generation. Wan 3.0 is API-only. - **Tencent's HunyuanVideo 1.5** (8.3 billion parameters) runs on consumer GPUs with about 14 GB of memory using offloading. Its license excludes the EU, UK and South Korea. - **MiniMax H3** released only its 33-billion-parameter base model, which outputs 768p. Its 2K stage is closed, and the license excludes the US, EU, UK and South Korea. FLUX 3's video model is API-only for now. See [LTX vs Wan](https://metavert.io/compare/ltx-vs-wan) for the two most usable open options. ### Provenance, Rights and Risk Realistic synthetic video raises the stakes of [deepfakes](https://metavert.io/deepfakes) and disclosure. Google watermarks every Gemini Omni output with SynthID, which it says has marked more than 100 billion images and videos. The [C2PA](https://metavert.io/c2pa) Content Credentials specification reached version 2.4 in April 2026. Under the [EU AI Act](https://metavert.io/eu-ai-act), Article 50's transparency obligations have applied since August 2, 2026. Providers of generative systems already on the market have until December 2, 2026 to meet the marking and detection duties. Platform rules are covered in [AI content labeling](https://metavert.io/ai-content-labeling). Training data remains contested. Seedance 2.0 drew a cease-and-desist from Disney within a day of its February 2026 release, along with objections from the Motion Picture Association, Paramount and Netflix. ### Where It's Headed The direction is toward longer, steerable, interactive generation. Clips are extended in conversation, characters persist across scenes, and the boundary with [world models](https://metavert.io/world-models) blurs. Generation is also moving inside editors: Adobe Premiere now offers Firefly and partner models directly in the timeline (see [AI video editing](https://metavert.io/ai-video-editing)). Whether to start from text or a frame is its own decision, covered in [text-to-video vs image-to-video](https://metavert.io/compare/text-to-video-vs-image-to-video). Because flagship models changed at least twice between March and October 2026, any specific ranking on this page should be read as a snapshot. ## Related Topics - [Generative Video](https://metavert.io/generative-video) — The broader field - [Text-to-Video vs Image-to-Video](https://metavert.io/compare/text-to-video-vs-image-to-video) — When to start from a frame - [Diffusion Transformers (DiT)](https://metavert.io/diffusion-transformers) — The architecture behind most video models - [Sora](https://metavert.io/sora) — OpenAI's discontinued video model - [Runway](https://metavert.io/runway) — A model maker with its own creative suite - [Kling](https://metavert.io/kling) — Kuaishou's video model family - [AI Content Labeling and Disclosure](https://metavert.io/ai-content-labeling) — Platform and EU rules for synthetic video - [World Models](https://metavert.io/world-models) — Where video generation is converging - [Veo](https://metavert.io/veo) — Google's Veo video models - [Gemini Omni](https://metavert.io/gemini-omni) — Google's newer video model, generally available since August 2026 - [Seedance](https://metavert.io/seedance) — ByteDance's video generation model - [Wan](https://metavert.io/wan) — Alibaba's Wan video models - [MiniMax](https://metavert.io/minimax) — Maker of the Hailuo models, including MiniMax H3 ## Further Reading - [Scalable Diffusion Models with Transformers](https://arxiv.org/abs/2212.09748) — Peebles and Xie, 2022; the DiT paper - [Gemini API pricing](https://ai.google.dev/gemini-api/docs/pricing) — Google; Veo 3.1 and Gemini Omni prices - [Gemini API deprecations](https://ai.google.dev/gemini-api/docs/deprecations) — Google; Veo shutdown dates - [OpenAI API deprecations](https://developers.openai.com/api/docs/deprecations) — OpenAI; the Sora API removal - [Text-to-video leaderboard](https://artificialanalysis.ai/video/leaderboard/text-to-video) — Artificial Analysis; blind-vote Elo rankings - [EU AI Act transparency obligations take effect](https://www.cooley.com/news/insight/2026/2026-08-03-eu-ai-act-transparency-obligations-take-effect-2-august-2026) — Cooley, August 2026