One AI video model made 4 in 10 videos

Section 01

The most used AI video model hasn't changed in 10 months

Kling 2.5 was the most used video model on the platform in November 2025 and it still was in August 2026. In between, its share of successful generations went from 30.1% up to 49.1% in January and back down to 31.4%, and it made 39.9% of the 108 million videos in the window.

Everything below it changed. Hailuo 2.3 and Seedance 1.0, second and third in November with 19.7% and 18.7%, are at 1.7% and 0% by August. Their places went to two models that didn’t exist on the platform in November: Seedance 2.0 at 18.9% and Kling 3.0 at 15.7%. Twice as many models were in use by August as in November, and the share of videos made by the top 5 didn’t budge: between 72.7% and 85.3% every month, and at its highest in July.

Section 02

New models that win take months to peak. The rest lose half their share in two

Two new models made it to the top of the chart in these 10 months, and neither did it fast. Seedance 2.0 took 16 weeks from its launch week to a peak of 26.7% of weekly videos. Kling 3.0 took at least 28 weeks to reach 16.3%, and it may not have peaked yet.

The other 14 launches we could watch for at least 8 weeks all peaked within 2 months, and none got past 7.5%. A fast peak signals a low ceiling. Falling is faster than rising: of the 14 models that ever held more than 5% of weekly videos, 8 have since lost half their peak share, in a median of 7.5 weeks. Seedance 1.0 went from 19.8% to under half of that in 7 weeks; Kling 2.1 halved from 11.3% in 11 weeks.

Section 03

Two companies made 8 in 10 AI videos

Group the models by who makes them and the picture gets simpler. Kuaishou, the company behind Kling, made 64.3% of all videos in the window. ByteDance, behind Seedance, made 17.1%. Together that's 81.5%.

The other 18.5% is split among MiniMax (Hailuo) at 7.5%, Google (Veo 3.1) at 5.5%, down from 8.2% of monthly videos in November to 3.5% in August, Alibaba (Wan and HappyHorse) at 3.2%, and 6 more makers plus a long tail, none above 1.5%. The balance is moving. Kuaishou peaked at 75.9% of monthly videos in Mar 2026 and was at 55.2% by August. ByteDance went from 7.1% in Mar to 28.0% in August, 4 times its share in 5 months.

Window

November 1, 2025 to August 31, 2026 (UTC), 10 complete calendar months, for monthly shares and the window totals. Weekly figures (adoption speed, half-life) use the 43 complete Monday-to-Sunday weeks inside that window, November 3, 2025 to August 30, 2026. Monthly volume is not published as a trend: generations follow platform traffic, and this report measures proportions.

Population

Successful video generations on the platform’s video generator flows (Freepik and Magnific combined), all surfaces except internal traffic. 108,302,430 generations in the monthly window; 107,731,361 in the 43-week window. Four internal services that run inside the video pipeline but don’t produce a video (an image-moderation pass above all, 1.97 million requests) were identified and removed; they had been counted as models in earlier drafts of this series. Every share in this report is over videos actually produced.

Definitions

A generation counts as successful when it returns status 200. A model is a maker’s model at a given version; resolution, duration, audio, and input mode are configurations of the same model. When the same model is served through more than one integration route, monthly shares count each route separately and maker shares add them up.

A model “launches” the first week it reaches 0.5% of weekly videos after the first week of the window. Half-life is the number of weeks between a model’s peak share and the first week it holds less than half of that peak, whatever the reason. Full definitions in the Report Glossary.

Limitations

Usage is not quality, and inside the interface it isn't purely preference either

The interface offers a default model, and 97% of videos come from the interface. In traffic with no interface, Kling 2.5 ranks third, not first. Nothing in this report scores what a model produces.

Half-life measures loss of share, whatever the cause

A model retired from the catalog counts the same as a model people stopped choosing; the data doesn't tell them apart. Three models peaked during a two-week swing in November 2025, when traffic moved from one model to another and back for reasons we're still confirming with the platform team; without them, the median half-life is 10 weeks rather than 7.5.

The newest models are cut off by the window

Seedance 2.5, Hailuo 3.0, and Wan 3.0 launched in the last 8 weeks and were still rising when the data ends. Their peaks and Kling 3.0's are lower bounds. The next quarterly edition closes them.

Only models available on the platform are measured

Shares reflect the platform's pricing and defaults. The low shares of Sora and Runway reflect what people choose at volume here, not a verdict on either model. We don't compare prices or free tiers, so this report doesn't say which generator is the best free one.

Reproducibility

Every number was independently recomputed by a second pass with its own queries before publication: the window total, the monthly ranking, the concentration table, the maker shares, and the exclusion of non-video services. Nine counter-queries, all reconciled to the published figures within rounding.

Cite & data

Cite this report

Citation

Magnific (2026). One AI video model made 4 in 10 videos. Creative Signals, September 2026. magnific.com/blog/ai-video-models

Short form for press

“One model, Kling 2.5, made 39.9% of the 108 million AI videos generated on Magnific and Freepik between November 2025 and August 2026, and was the most used model in all 10 months.”