Chinese labs hold nine of the ten arena video rows
As of mid-August 2026, Chinese developers occupy nine of the top ten text-to-video rows on Artificial Analysis; Google is the exception. An Elo table is not a job table — pick Seedance 2.5, Kling 3 Turbo, Happy Horse, Wan, Vidu or Hailuo by the published row.
As of mid-August 2026, Chinese developers occupy nine of the top ten text-to-video rows on Artificial Analysis; Google is the exception. An Elo table is not a job table — pick Seedance 2.5, Kling 3 Turbo, Happy Horse, Wan, Vidu or Hailuo by the published row.
The Japan Times reprint of Catherine Thorbecke's Bloomberg commentary put the split in one sentence: outside Alphabet's Google, nearly all of the leading video-generation models are Chinese, nine of the top ten text-to-video systems on Artificial Analysis. ByteDance and MiniMax had just shipped rival updates. Happy Horse had rattled the industry unnamed, then been claimed as Alibaba. Kuaishou's Kling and ShengShu's Vidu sit in the same crowded roster. OpenAI and Anthropic, the piece notes, deprioritized the field. That is a snapshot of a public board. It is not a shopping list.
What the mid-August board actually measured
An arena Elo moves when two clips from the same prompt sit next to each other and a voter picks one. The clip is short. The voter is not holding your pack shot, not checking a spoken line, not paying your credit bill. They are scoring whether eight-ish seconds looked finished.
Nine Chinese rows and one Google row tells you who showed up to that contest. It does not tell you which published duration, resolution, credit sticker, or audio flag matches the job on your desk. Treat the Japan Times piece as weather, the way Arena Elo is not the buying guide already asks you to. Cite mid-August 2026 as a snapshot.
The labs in that screenshot ship more than one row. Kling Turbo is not Kling Pro. Hailuo Fast is not MiniMax H3. The board collapses those differences into a flag. Google as the exception is also collapsed: Veo 3.1 is 4 / 6 / 8 seconds, a 4K tier, 40 credits, native audio. "Google" is not a model. Neither is "Chinese video."
An Elo table is not a job table
Your brief is almost never "win a blind vote on a paragraph." It is a long hold, a silent fast plate, a 1080p clip in a 3–15 second window, or a draft from a still at 6s or 10s. Those jobs have different winners, including models that were never going to top a text-to-video arena because they are not that contest.
| The actual job | Published row | What Elo is silent on |
|---|---|---|
| One ~30s hold at 720p | Seedance 2.5: 4s–30s, 47 credits, native audio | Whether second 22 still matches second 2 |
| Fast silent 1080p motion | Kling 3 Turbo Text to Video: 3s–15s, 12 credits discounted, audio false |
Whether the file has a stem |
| 3s–15s 1080p from a paragraph | Happy Horse 1.0 Text to Video: 28 credits | Whether the mystery-lab story still matters |
| 2s–15s 1080p from a paragraph | Wan 2.7 Text to Video: 15 credits | Driving audio is not a native-audio promise |
| 5s / 10s / 15s 1080p with audio | Vidu Q3 Video: 16 credits, native audio | Three duration steps, not a slider |
| Draft from a still, 6s or 10s | Hailuo 2.3 Fast: 28 credits discounted, image-to-video | A T2V board never saw your still |
Read the row: duration enum, resolution ceiling, credit sticker, discounted flag, audio flag. If audio is false or unset, do not write dialogue into the prompt and wait for a voice file. If the job starts with an approved frame, the T2V leaderboard is a spectator sport — open image-to-video and pick a still-required row.
Pick the published row, not the nationality
The commentary's useful list is the labs, not a crown. ByteDance, MiniMax, Alibaba (Wan and Happy Horse), Kuaishou, ShengShu. Each name is several SKUs. The AI video generator is where you press generate. The press clipping is not.
Seedance 2.5 is the long single-shot contract: up to 30 seconds, 720p, 47 credits, native audio. Pick it because 30 seconds in one pass is the product, not because ByteDance "won China." Kling 3 Turbo is the cheap silent plate: 3s–15s, 1080p, 12 credits discounted, audio false. Quoted lines do not become a soundtrack. Happy Horse 1.0 is Alibaba's named 1080p T2V row, 3s–15s, 28 credits — the mystery-lab story is already spent. Wan 2.7 is the other Alibaba landing: 2s–15s, 1080p, 15 credits; prompt rewriting is listed, native audio is not. Vidu Q3 Video is ShengShu at 5s, 10s, or 15s, 1080p, 16 credits, native audio. Hailuo is MiniMax's still-to-motion family, not the T2V arena: Fast is 6s or 10s at 28 credits discounted; Standard is the same durations at 28 credits with audio false.
Google remains the exception on the board and a different job in production. Veo 3.1 still wins a talking 4 / 6 / 8 second clip at 4K. An arena week that puts Omni-shaped Google above Veo does not move the mouth.
How to use a 9-of-10 screenshot without becoming it
Name the job in one line before you open Artificial Analysis. Filter to inputs the job actually has. Read credits for a complete generate: forty-seven credits of Seedance 2.5 is a 30-second hold; twelve discounted credits of Kling Turbo is a silent 1080p plate. Those are not interchangeable "Chinese video." Keep Elo as a tie-break inside one row — two models that both fit duration, credits, and audio flag — then the prettier 8-second vote can decide. Not before.
Do not migrate off Veo because nine of ten rows are Chinese. Do not migrate onto Happy Horse because Alibaba's mystery model "won." Open the catalog. Pick the row. The screenshot can stay in the Slack thread.
FAQ
Is Google really the only non-Chinese lab in the top ten?
As of the mid-August 2026 snapshot the Japan Times carried, yes: nine of the top ten text-to-video systems on Artificial Analysis were Chinese, with Google the exception. That sentence will rot when the board moves. The jobs will not.
Should I always pick the arena leader?
No. The leader won a short pairwise vote. If your brief is a 30-second identity hold, Seedance 2.5 is the published row even when it is not the screenshot. If your brief is a silent 5-second hook, Kling 3 Turbo is the cheap plate. Elo can break a tie inside that shortlist. It cannot pick the shortlist.
Does this mean I should abandon Veo?
No. Veo is the model for always-on dialogue and a 4K talking clip at 4 / 6 / 8 seconds. A lower arena rank is not a reason to abandon the mouth. Google being the exception on a T2V board is a fact about who showed up to that contest, not a fact about your talking-ad job.
Will the 9-of-10 split last?
Do not plan a quarter on it. ByteDance, MiniMax, Alibaba, Kuaishou and ShengShu will keep shipping. Google will keep shipping. Arena weeks are weather. Pick Seedance, Kling Turbo, Happy Horse, Wan, Vidu or Hailuo by the published duration, credits, and audio flag — the same way you would if the screenshot said the opposite.