Compare AI video models, task by task

There is no single winner for every workflow. Text-to-video, image-to-video, native audio, duration, speed, price, API access, and local execution reward different models. The two LMArena rankings therefore remain separate.

It compares 24 relevant exact model versions across 276 pair pages. Published blind-preference ratings remain separate from pricing and technical specifications.

Text-to-video starts from a written prompt. Image-to-video animates a supplied still image. A model may rank differently because the two tasks demand different types of control and consistency.

No. Ratings are comparable only within the same published leaderboard and dataset revision. A text-to-video rating must not be ranked against an image-to-video rating.

A price appears only when a traceable public rate can be tied to the exact model configuration. Credit packages, subscriptions, and reseller rates are not converted into a single number without a reliable basis.

23 of 24 exact model versions currently have a traceable first-output run. Every included model receives the identical archived prompt in one request. Playback is normalized to the first five seconds, 16:9, 1,280 × 720 pixels, 30 fps, and muted audio without cropping. Provider-side randomness and prompt processing remain model characteristics.

The stored LMArena data were retrieved on 2026-09-22 from an immutable repository revision. Model specifications and prices link to their sources and carry individual verification dates. Sora's retirement was checked on September 28, 2026.