AI Video Arena: Wan 3.0, Kling, Seedance, Veo, H3, and Runway Compared
I tested 6 AI video models on 5 identical scenes: Wan 3.0, MiniMax H3 Max, Kling 3.0, Seedance 2.5, Veo 3.1 Lite and Runway Gen-4.5. Side-by-side clips, scores and cost per clip.
👁 140 reads
Wan 3.0 won my AI Video Arena with an average of 8.8/10 across five identical scenes, ahead of MiniMax H3 Max at 8.4. They were the only two of six models that wrote the Polish words “ŻÓŁTA ŁÓDŹ” (“yellow boat”) on a chalkboard without broken letters, and Wan cost just $0.21 per 5-second clip. Veo 3.1 Lite (4.8) barely moved the camera and wrote Cyrillic, while Runway Gen-4.5 (3.9) lost a full scene because its API returns no audio. The round ran on October 8, 2026, through OpenRouter, for $8.54 in total.
For a broader comparison of AI video generators, with my test results, USD prices and free plans, see Best AI Video Generator 2026: Test Results and USD Prices.
Wan 3.0 leads on four of five scenes
Each scene is scored from 0 to 10. The average covers all five scenes, so a model without audio output gets 0 for the speech scene. The speech clips come from my September 28 round with the same prompt; the other 25 clips are from October 8.
| Model | Camera movement | Character speaking Polish | Text on chalkboard | Water physics | Character consistency across 2 cuts | Average |
|---|---|---|---|---|---|---|
| Wan 3.0 | 8.5 | 7.5 | 10.0 | 9.0 | 9.0 | 8.8 |
| MiniMax H3 Max | 8.0 | 9.0 | 10.0 | 7.5 | 7.5 | 8.4 |
| Kling 3.0 Standard | 8.5 | 5.5 | 4.0 | 7.5 | 7.5 | 6.6 |
| Seedance 2.5 | 6.0 | 4.0 | 5.0 | 7.0 | 6.0 | 5.6 |
| Veo 3.1 Lite | 3.0 | 7.5 | 2.0 | 5.0 | 6.5 | 4.8 |
| Runway Gen-4.5 | 6.5 | 0.0 | 2.0 | 4.0 | 7.0 | 3.9 |
Costs and generation parameters for this round:
| Model | Average without speech | Cost per clip | Duration | Resolution | API Audio |
|---|---|---|---|---|---|
| Wan 3.0 | 9.1 | $0.21 | 5 s | 480p | yes |
| MiniMax H3 Max | 8.2 | $0.25 | 5 s | 480p | no |
| Kling 3.0 Standard | 6.9 | $0.42 | 5 s | 720p (1080p file) | yes |
| Seedance 2.5 | 6.0 | $0.52 | 5 s | 480p | yes |
| Veo 3.1 Lite | 4.1 | $0.18 | 6 s | 720p | yes |
| Runway Gen-4.5 | 4.9 | $0.48 | 4 s | 720p | no |
What separated the models:
- On-screen text rendering created the widest separation. Every model wrote the plain first line “ARENA WIDEO 2026”, but only Wan 3.0 and H3 Max got the accented second line right.
- Camera control: Kling, Wan, and H3 Max executed a clean 180-degree orbit around the motorcycle. Veo 3.1 Lite barely moved the camera at all.
- Fluid simulation: Only Wan depicted the full sequence in logical order - glass filling up, overflowing, and forming a puddle on the table surface.
- Pricing vs. quality: The best clips were not the most expensive. Wan 3.0 at 480p cost $0.21 for 5 seconds, whereas Seedance 2.5 at the same resolution cost $0.52.
Scene 1: camera movement
Evaluation criteria: smoothness of the 180-degree orbit, stability of the motorcycle and background, realistic puddle reflections.
Prompt used for all models
A smooth 180-degree camera orbit around a vintage red motorcycle parked on a wet cobblestone street in an old European town at blue hour. The camera circles steadily from the front of the motorcycle to its back at a constant speed. Street lamps reflect in the puddles, light drizzle. The motorcycle keeps its exact shape and details. Realistic cinematic footage, real-time motion, no text, no logos.
Scene 2: a character speaking Polish
Evaluation criteria: intelligibility of the spoken line (transcribed via faster-whisper large-v3, measured by Word Error Rate - WER), lip-sync and accent. I use a Polish tongue twister because its consonant clusters expose weak speech models quickly.
The clips come from my September 28, 2026 round with the same prompt; only Veo 3.1 Lite got a new generation on October 8. Full transcripts and the results for full Veo 3.1, Gemini Omni Flash and Grok Imagine are in my AI video model ranking.
Prompt used for all models (includes the spoken Polish tongue twister)
Medium close-up of a woman in her thirties sitting at a kitchen table, looking straight into the camera. She says clearly and calmly in Polish: "W Szczebrzeszynie chrząszcz brzmi w trzcinie, a ja płacę trzydzieści dwa złote." Natural daylight from a window, static camera, realistic, no subtitles, no on-screen text, no music.
Translation of the spoken sentence: “In Szczebrzeszyn a beetle buzzes in the reeds, and I am paying thirty-two zlotys.”
Scene 3: only two models spelled the chalkboard text correctly
Evaluation criteria: exact reproduction of both lines, with the Polish letters Ż, Ó, Ł and Ź as the hard part.
Prompt used for all models (contains Polish characters)
Close-up of a green school chalkboard in a sunlit classroom. A hand finishes writing with white chalk. The chalkboard clearly shows exactly two lines of text in capital letters: first line "ARENA WIDEO 2026", second line "ŻÓŁTA ŁÓDŹ". Static camera, realistic, chalk dust in the light, no other text.
Translation of the chalkboard text: “VIDEO ARENA 2026” and “YELLOW BOAT”.
Scene 4: fluid dynamics and water physics
Evaluation criteria: continuous stream, water splashes, rising liquid level, realistic overflow onto the table, and absence of disappearing fluid.
Prompt used for all models
Static close-up on a wooden kitchen table in daylight. Water is poured from a glass pitcher into an empty clear drinking glass. The stream splashes, bubbles rise, the water level rises steadily until the glass overflows and water spills onto the table and spreads into a puddle. Real-time speed, no slow motion, realistic physics, no text.
Scene 5: character consistency across hard cuts
Evaluation criteria: matching facial features, hair, glasses, raincoat, and backpack across a hard cut.
Prompt used for all models
Two shots separated by one hard cut. Shot 1: wide shot of a young woman with short red hair, round black glasses, a yellow raincoat and a black backpack walking along a rainy city street at night with neon reflections on the wet pavement. Hard cut. Shot 2: close-up of the same woman with the same short red hair, the same round black glasses and the same yellow raincoat as she stops under a shop awning and looks up at the rain. Realistic, cinematic, no text.
Test methodology
- Every generator received the same English prompt (only the chalkboard text and the spoken line are in Polish), 16:9 aspect ratio, pure text-to-video (no init images), and single-pass generation - no cherry-picking or rerolling.
- Target length was set to 5 seconds. Duration variances reflect model constraints: Veo 3.1 Lite requires 4, 6, or 8 seconds (6s selected); Runway Gen-4.5 was set to 4s to stay within budget constraints.
- Resolutions: Seedance 2.5, Wan 3.0, and H3 Max at 480p; Kling, Veo, and Runway at 720p (the lowest native options available). Kling 3.0 Standard delivers a 1080p container regardless.
- Scoring: I watch each generation end to end and compare 4 extracted frames (at 8%, 35%, 62%, and 92% timestamps) side-by-side. Text accuracy is validated on the final frame at native resolution.
- Speech analysis: Transcribed via faster-whisper large-v3 with language set to Polish, measuring Word Error Rate (WER) after converting numeric values to full words.
- Cost tracking: taken directly from the
usage.costparameter returned by OpenRouter API, recorded in USD without sales tax. - Media playback: All video assets are transcoded to H.264 (max width 1280px) and configured to load on demand.
Models left out of the October 8 round
- Kling 4.0: on October 8 it was not callable through any API, so Kling appears here as 3.0 Standard. Kling 4.0 opened to all Kling AI users on October 15, 2026, but only in the app, on kling.ai and through the official MCP and CLI; it joins the Arena when a per-second API is available. Details: Kling 4.0 is now available.
- Sora 2: Unavailable. The consumer interface ceased operations on April 26, 2026, and the Sora 2 API was shut down on September 24, 2026.
- Full Veo 3.1 and Gemini Omni Flash: Full-tier Veo costs $0.40 per second with audio, and Omni Flash is not routed through OpenRouter. Both were tested in my September 28 round in the AI video ranking.
- Kling 3.0 Pro: Excluded for budget reasons ($0.56 per 5 seconds without audio). My image-to-video test of Kling 3.0 Pro is in the Kling AI guide.
Later rounds on the same prompts
- HappyHorse 1.1 (Alibaba), October 11: 6.1/10 on the four non-speech scenes. Clips: HappyHorse vs Wan 3.0, Seedance and MiniMax; model page: HappyHorse.
- FLUX 3 Video, October 11: 8.9/10, with correct Polish letters and speech (FLUX 3 Video).
- Vidu Q4 Preview has no text-to-video mode, so it needs starting frames before it can enter (Vidu Q4 Preview).
Frequently asked questions
What is the cheapest AI video model in this test?
Veo 3.1 Lite had the lowest price per clip, $0.18 for 6 seconds, but finished second to last. Wan 3.0 is the better buy: $0.21 for 5 seconds at 480p and the top average.
Can I use Kling 4.0 through an API?
Not at launch. On October 15, 2026, Kling’s developer page still said “Kling 4.0 API Coming Soon,” and only the official MCP server and CLI worked. My check of seven providers is in Kling 4.0 API: where it works.
Can you still use OpenAI Sora?
No. The Sora web app was discontinued on April 26, 2026, and the Sora 2 API was decommissioned on September 24, 2026. OpenAI announced the two-stage shutdown on March 24, 2026.
Sources
- API pricing and endpoints: OpenRouter Video Models Directory (verified October 8, 2026).
- Kling 4.0 rollout: Kling Release Notes.
- Sora API deprecation: OpenAI API Deprecations Documentation.
- Sora shutdown dates: The Decoder - OpenAI sets two-stage Sora shutdown with app closing April 2026 and API following in September.
- Speech transcripts: faster-whisper large-v3, my September 28 and October 8, 2026 runs.
- Checked on October 29, 2026.
