Which AI Video Models Do People Actually Use? 2026 Data From 4,069 Jobs

Short answer: Seedance v2 text-to-video was the most used AI video model on Oakgen this year. The Oakgen AI Video Index 2026, based on 4,069 video generations from January 10 to November 10, 2026, found it ran 22.2% of all jobs. No other single model passed 8%.

The bigger finding is about how people start. 64.5% of jobs began from an image, not a text prompt. Most people already have a picture of the product, the person or the scene. They want it to move.

The full table with sample sizes is on the AI video statistics hub.

Which AI video models do people use most?

Share of all video jobs, January 10 to November 10, 2026 (n = 4,069).

• Seedance v2 text-to-video ran 22.2% of video jobs.
• Trend videos ran 14.5% of video jobs.
• Seedance v2 image-to-video ran 7.2% of video jobs.
• Kling v3 Pro image-to-video ran 6.1% of video jobs.
• MiniMax H3 reference-to-video ran 5.2% of video jobs.
• Grok image-to-video ran 5.1% of video jobs.
• Kling v3 Pro text-to-video ran 5.1% of video jobs.
• Kling v4 image-to-video ran 4.2% of video jobs.
• The video upscaler ran 2.9% of video jobs.
• Grok text-to-video ran 2.9% of video jobs.

The top ten add up to about 75%. The rest is spread across models with fewer than 100 jobs each.

How reliable is each AI video model?

Completion rate is the share of a model's jobs that finished.

Most single models finished more than 94% of jobs. The two outliers are not normal generators. The upscaler works on an existing video, and trends put your photo into a fixed scene, which trips more filters. We explain that in why AI video trends get blocked.

Do people use image-to-video or text-to-video more?

Mode is set by the inputs. Any image or reference image counts as image-to-video. Any input video counts as video-to-video. The rest is text-to-video.

• Image-to-video was 64.5% of video jobs (n = 4,069).
• Text-to-video was 28.5% of video jobs (n = 4,069).
• Video-to-video was 7.0% of video jobs (n = 4,069).

Image-to-video wins because it gives you control. You choose the face, the product and the framing. The model only adds motion. A text prompt leaves all of that to chance.

Is text-to-video more reliable than image-to-video?

On average, yes.

• Text-to-video completed 93.4% of the time (n = 1,159).
• Image-to-video completed 87.3% of the time (n = 2,625).
• Video-to-video completed 77.9% of the time (n = 285).

Look at single models and the gap closes. Seedance v2 image-to-video finished 97.9% of jobs and Grok image-to-video finished 97.6%. So the lower average comes from a few groups of jobs, not from using an image as such. The data does not say which images failed, and we do not look at them.

So the lesson is not "avoid images". It is "watch what is in the image". A photo you own is a safer input than a screenshot of a film or a famous face, which is the kind of thing content and copyright filters exist to catch. Our post on how often AI video generations fail covers the error codes.

How should you choose an AI video model?

Popularity is a starting point, not an answer. Here is how we would pick.

1. Start from an image if you have one. It is what most people do, and it gives you control.
2. Draft on Grok or MiniMax. They were the cheapest per clip and finished more than 96% of jobs. See AI video cost per finished clip.
3. Finish on Kling or Seedance. They carry most of the volume for longer, polished shots.
4. Use the upscaler last. It had the lowest completion rate of the regular tools, so run it only on a clip you will keep.

For a job-by-job guide, read the best AI video generators by use case.

How does this fit the wider market?

AI video use is spreading fast. Wyzowl's State of Video Marketing 2026 survey (266 respondents, late 2025 fieldwork) found 63% of video marketers have used AI to create or edit video, up from 51% a year earlier (Wyzowl, 2026). Kuaishou said Kling served more than 60 million creators by December 2025 (Kuaishou Technology, January 13, 2026).

Those numbers say how many people use AI video. They do not say which models people pick when they can choose. That is what a multi-model platform can show.

How we measured this

• Dataset. The Oakgen AI Video Index 2026 covers every video job on Oakgen from January 10 to November 10, 2026 (n = 4,069).
• Model. Each job is grouped by the model stored on it. Trend videos are grouped as one pipeline.
• Completion. A model's completion rate is its completed jobs divided by all its jobs.
• Privacy. Aggregate counts only. No prompts, images or user-level data.
• Suppression. Models with fewer than 100 jobs are left out.

What are the limits of this data?

• Single platform. Share reflects Oakgen's model menu, defaults and prices, not the whole market.
• Availability changed. Some models launched during the period, so they had fewer months to collect jobs.
• Small samples. Several models sit just above 100 jobs. A few failures move their rate by a point.
• Uneven months. October 2026 alone had 1,029 video jobs, a quarter of the total, so recent choices weigh more.

See all cuts on the AI video statistics hub.

Frequently asked questions

What is the most used AI video model on Oakgen?

Seedance v2 text-to-video. It ran 22.2% of the 4,069 video jobs in the Oakgen AI Video Index 2026 (January 10 to November 10, 2026). Trend videos were second at 14.5%.

Which AI video model fails least?

Among models with at least 100 jobs, Grok text-to-video (98.3%, n = 116) and Kling v3 Pro text-to-video (98.1%, n = 207) had the highest completion rates, followed by Seedance v2 image-to-video at 97.9% (n = 291).

Does popularity mean a model is best?

No. Share reflects defaults, price, features and what was available during the period. Use it to see what others chose, then pick by the job you need done.

Sources

1. Oakgen AI Video Statistics 2026
2. Wyzowl: The State of Video Marketing 2026
3. Kuaishou Technology: Kling AI Annualized Revenue Run Rate Hits USD240 Million in December 2025 (2026-01-13)