Text-to-Video AI Rankings: Wan 3.0 and Gemini Omni Lead Post-Sora

A new top tier

The leaders in AI video generation changed over the summer of 2026. By September, the top of the major leaderboards was held by Alibaba's Wan 3.0, Google's Gemini Omni Flash and several other Chinese models. Last year's best-known names had slipped down the rankings. One widely shared September ranking put Wan 3.0 first on the Artificial Analysis text-to-video leaderboard, with Gemini Omni Flash second, followed by MiniMax H3, ByteDance's Seedance 2.5 and Kuaishou's Kling 3.0 Turbo.21 Google's Veo 3.1 now sits further down that list. The same ranking describes it as the safest and most widely available Western option, not the quality leader.21

OpenAI's withdrawal is the biggest part of the story. Sora's web and app experience closed on April 26, and the API was scheduled to shut down on September 24.4 One comparison says the most famous Western video product has been switched off completely. It describes the top of the with-audio board as split between Google and a group of models from Alibaba, ByteDance, MiniMax and Kuaishou.13 Some consumer roundups still list Sora as a top pick for photorealism, even while noting that it was discontinued.27 That shows how far behind some buyer's guides have fallen.

The two leaderboards don't agree

People talk about "the leaderboard," but there are actually two influential public rankings, and they disagree at the top. Artificial Analysis rebuilt its text-to-video board as version 2.0, with an Elo scale fixed so that Kling 3.0 1080p (Pro) scores exactly 1,000. Scores from earlier in the year can't be compared with current ones.11 On that board, Wan 3.0 ranks first both with audio (1,156) and without (1,128). Seedance 2.5 is third on the with-audio board at 1,143, and MiniMax H3 is fourth at 1,137.1

The Text-to-Video Arena tells a different story. Its September 21 snapshot drew on about 718,000 votes across 48 models. It placed Google's gemini-omni-1.1-flash first at 1,516 and the original gemini-omni-flash second at 1,513. Black Forest Labs' flux-3-video and SpaceXAI's grok-imagine-video-1.5-agent came next.28 On that board, Wan 3.0 is sixth at 1,476. Its rank spread runs from three to nine, and it has only about 2,600 votes.28

Our reading: the evidence supports a crowded top tier, not a single winner. The Arena's top four entries all have rank spreads that start at first place, so the statistical uncertainty overlaps.28 Rankings that crown one champion are claiming more precision than the data allows. One builder-focused guide reaches the same conclusion and advises choosing models by output shape first and leaderboard position second.11

What happened to last year's leaders

The drop for established names is striking. Runway Gen-4.5 was first on Artificial Analysis when it launched in December 2025, with an Elo of 1,247 on the old scale.11 Runway hasn't released a successor, and one analysis now places Gen-4.5 at 23rd on the no-audio board.11 On the Arena it is 27th with a score of 1,225. Every Veo 3.1 variant ranks between 12th and 15th, and Sora 2 Pro is 11th.28 An earlier snapshot from April 30 shows how quickly things moved. Back then, Alibaba's HappyHorse-1.0 led at 1,368, Runway Gen-4.5 was eighth and Veo 3.1 was ninth.8

Review sites and leaderboards measure different things, which explains much of the gap. Some reviewers still name Runway the best overall tool, citing its Gen-4.5 output plus a full editing suite.22 Others prefer Runway for most creative teams because it combines generation, reference control and in-app editing.23 Another rates Google Flow with Veo 3.1 highest at 9.1 out of 10, with Runway second at 8.9.19 These are product judgments about workflow and tooling. Arena Elo only records which clip people preferred in blind side-by-side votes. A podcast analysis puts it this way: Runway is a good generator attached to the best editing model, Aleph, and an API that now resells other labs' models.18

Native audio is now standard

The biggest technical change of the year is in AI voice and sound, not visuals. In May 2025, Veo 3 was the only mainstream video model that produced sound with its pictures. By July 2026, Wan, LTX, PixVerse, Seedance, Kling, HappyHorse, MiniMax and Black Forest Labs had all released models that generate audio and video in one pass.2 An industry review from July reached the same conclusion: synchronized audio is now a basic requirement, and a top lab launching a silent model would look like a step backward.8 A February analysis made a similar point earlier in the year, when four of the six major models already had native audio.10

Competition has moved to the quality and type of sound. HappyHorse leads dialogue support with seven languages, Kling 3.0 handles five, and MiniMax H3 is the only model claiming stereo output.2 Kling 3.0 generates speech in Chinese, English, Japanese, Korean and Spanish.1 MiniMax H3 offers 32kHz stereo and dialogue in 11 languages.11 FLUX 3 claims lip-synced dialogue in more than 14 languages.3 Google itself admits that short speech segments are still a weak spot for Veo.2

Runway's audio is a point where coverage contradicts itself. One guide says Gen-4.5 now generates native audio alongside video.14 Others list it as having no native dialogue23, or as a silent model on current boards.11 Another says audio comes through its Aleph integration.25 The most consistent reading is that Runway's sound comes from separate tooling, not from the core generation pass. That matters a lot now that rivals bundle audio by default.

Longer clips and a new class of multimodal models

The second shift is length and reference handling. Wan 3.0 entered beta on August 6 and became generally available on Alibaba Cloud on August 24. It produces single-pass clips up to 30 seconds long at up to 1080p and accepts up to 20 reference inputs.1 Seedance 2.5 launched July 31 with 30-second audio-video clips and multiple rounds of extension.4 One review says it accepts up to 50 image, video and audio references.3 Meanwhile, Kuaishou opened Kling 4.0 Flash to a limited early-access group on September 28, promising clips of 3 to 30 seconds, up to 4K, stereo audio and 10-bit HDR. The full release is expected in October.1

These releases are increasingly single models that handle text, images, video and audio together. MiniMax describes H3 as working with text, image, video and audio context in one model.4 FLUX 3 combines image, video and audio generation in one architecture.3 That links video directly to developments in AI image and voice models.

Open weights: a dispute worth settling

The biggest factual disagreement is about openness. One September ranking calls Wan 3.0 an open-weight release you can use without a closed platform.21 Another comparison says it is API-only and that Alibaba's open Wan line stops at version 2.2.13 The Arena lists wan3.0 as proprietary.28 The evidence favors the second view: Wan 3.0 should be treated as a closed, hosted model.

The strongest open-weight option is MiniMax H3. Its weights were released August 3, and it ranks fourth on the with-audio board.1 There are caveats. The free weights exclude the US, EU, UK and South Korea.11 The open base checkpoint also doesn't include hosted stages such as the 2K regeneration step.4 LTX-2.5 from Lightricks arrived August 11.1 Kandinsky 6.0 Video was released under the MIT license on October 6.1 Open models have mostly caught up on features like audio, but they still trail closed models on quality.8

Why it matters

There are three practical conclusions. First, a ranking older than about six weeks is unreliable, because the scoring scales were reset this year and top positions have changed several times.115 Second, the strongest models now come mostly from China. That brings access limits: Seedance 2.5's API is not offered in the US.11 Some buyers also have concerns about using Chinese vendors.13 Third, price matters as much as quality. Gemini Omni Flash is listed at $0.10 per second21, and Veo 3.1's API pricing ranges from $0.05 to $0.60 per second.11

For buyers, the useful question is no longer which model is best overall. It's which model fits a specific job: clip length, dialogue languages, budget and where the work will be distributed. Sora's exit made one thing clear. In this field, even a top position can disappear within a year.