Skip to content
trending

State of AI

SkyReels V4: From No. 1 to No. 9 in Seven Months

SkyReels V4 led Artificial Analysis’s video ranking in March 2026. By October, eight newer models had overtaken it.

By 4 min read

👁 114 reads

On March 18, 2026, Artificial Analysis named SkyReels V4, from Chinese company Skywork AI, the new leader in text-to-video with audio, ahead of Kling 3.0 and Veo 3.1. Less than seven months later, the same model is No. 9. It has not got worse - eight newer models have overtaken it, all released since March. It is a clear example of how quickly video generators age, and why a ranking from last quarter can already be out of date. Read more about the model itself: SkyReels V4.

March: an unexpected leader

Skywork AI, a subsidiary of Chinese company Kunlun Tech, was known for open models for avatars and talking characters. On February 25, 2026, it published a report on SkyReels V4 - the first full video model with audio that can also edit footage: replacing objects and backgrounds, changing the time of day, and transferring motion from one video to a character in a photo. In the report, the model ranked No. 3 in the AA Arena, behind Veo 3.1 and a model from Grok. Three weeks later, when its API launched, it moved into first place for text-to-video with audio.

According to Skywork’s report, it also ranked No. 2 in text-to-video without audio, No. 4 in image-to-video with audio, and No. 7 without audio.

October: No. 9

Here is the current Artificial Analysis text-to-video table (version v2.0):

Rank Model Elo Release date according to AA
1 Wan 3.0 1156 August 2026
2 Utopai X 1149 September 2026
3 Seedance 2.5 1143 July 2026
4 MiniMax H3 (768p) 1137 July 2026
5 MiniMax H3 Max 1130 August 2026
6 FLUX 3 1128 July 2026
7 Seedance 2.0 1116 March 2026
8 Gemini Omni Flash 1.1 1113 August 2026
9 SkyReels V4 1074 March 2026
16 Kling 3.0 1080p Pro 1000 February 2026
19 Veo 3.1 962 January 2026

Source: Artificial Analysis AA-Video-T2V v2.0, October 11, 2026. In image-to-video, SkyReels V4 ranks 17th out of 36, with 1070 Elo.

Dealat Kinetyra
Synthesia Starter - 12 Months
$79 instead of $216 -63%
See the deal →
Release month of models in the top nine (AA, text-to-video)
March 20262
July 20263
August 20263
September 20261

Only two of the top nine models are from March: Seedance 2.0 and SkyReels V4. The other seven launched in July, August, and September.

Why the rankings change so quickly

  • Release pace: In summer 2026, new video models near the top appeared every few weeks - MiniMax H3 in late July, FLUX 3 Video on August 4, Wan 3.0 in mid-August, Gemini Omni Flash 1.1 on August 27, and Utopai X on September 30.
  • Fine-tuning open models: Two top-five models were built from the open MiniMax H3 base: H3 Max (fal.ai) and Utopai X (a film studio). New players no longer need to train everything from scratch.
  • Relative rankings: Elo measures how often a model beats others. When stronger models enter the Arena, an older model loses points, even if it generates exactly what it did in March.
  • Changes to the leaderboard: AA has reworked its rankings (text-to-video is now version v2.0), so scores from different months are not directly comparable - ranks are.

What SkyReels V4 still does well

A drop in the rankings does not mean the model is useless. SkyReels V4 still ranks ahead of Kling 3.0 and Veo 3.1, the models it beat in March. Its strongest capability is not measured by the text-to-video and image-to-video tables at all: it can edit existing footage. In Skywork’s report, it replaces a bison with a moose in a selected area, changes the background behind a person to the Amalfi Coast, turns day into night, and transfers motion from one video to a character in a photo. I collected five examples with prompts in What SkyReels V4 Can Do.

SkyReels V4 transfers motion from a video to a character in a photo
“Animate the person in @image_1 using the movements from @video_1.” Translation of the prompt used in the original Polish test material. Frames from Skywork AI’s paper (arXiv 2602.21818, CC BY 4.0).

The price has not changed: the official API costs $0.35 per second for 1080p ($21 per minute according to AA), while Fast at 480p starts at $0.08. That is roughly the same as Seedance 2.0 and more than Wan 3.0 ($12) or MiniMax H3 ($4.80) - models that now rank higher.

What to consider when choosing a video generator

  • Check the ranking date. A comparison from three months ago may have a different leader and a different top five.
  • Avoid an annual plan tied to one model. March’s leader is now in the middle of the pack; pay-per-second APIs and platforms offering multiple models give you more flexibility.
  • Choose for the task, not the rank. SkyReels V4 can still make sense for editing footage. In my Polish-language tests, Wan 3.0 and FLUX 3 Video handled Polish dialogue better.

I track the current leaders using the same prompts in the Video Arena, and you can find a broader overview in the AI video generator rankings. For a comparison of models from the same March wave, see SkyReels V4 vs. Seedance 2.0 vs. Kling 3.0.

// read next

More on this from QUANTSLANT

Piotr Olszewski

Editor and publisher · Quantslant

I edit and publish Quantslant. AI tools help with research, drafts and translation; choosing topics, running the tests, checking facts and the decision to publish are mine, and so is the responsibility for what appears here.

read nextbenchmarksDecision Models vs Small LLMs: Cost, Speed and AccuracyArtificial Analysis benchmarks show decision models can be much faster, but Claude Haiku 5.5 (low) scores highest. Compare accuracy, latency and costs.

Most read

  1. 1Best AI Video Generator Apps for Your Phone (2026)
  2. 2Claude Opus 5.5 Motion Graphics: 60s Showreel from One Prompt
  3. 3Wan 2.2 on RTX 4090: timings, VRAM and four test clips
// site map

This page in other languages9

Quantslant in numbers 1,427posts in archive26news from the last 7 days18video models in the observatory
🎬 AI Video Benchmarks: Higgsfield DoP vs OpenAI Sora • Higgsfield vs Runway Gen-3 • AI Video Credit Calculator
× ‹ enlarge ›