The Video Leaderboard's Number One Is a Talking Head

The number one video model on OpenRouter for the week ending October 28 is not a cinema model. It is HeyGen Video, at 64,000 requests.

Second is Veo 3.1 Lite at 42,000. Third is Wan 3.0 at 39,000. Fourth is Seedance 2.0 Fast at 35,000, up 89% in a week. Fifth is Seedance 2.5 at 32,000. Then Grok Imagine Video 1.5 Lite at 31,000, Seedance 2.0 at 19,000, Veo 3.1 Fast at 17,000, Avatar IV at 13,000, and Kling Video v3.0 Pro at 13,000.

Two of the ten are the same company's avatar products. HeyGen holds first place and ninth. And in the author market share table for the week beginning October 19, ByteDance leads at 24.1%, Alibaba is second at 19.6%, HeyGen is third at 17.3%, Google is fourth at 15.3%.

Add ByteDance, Alibaba, Kwaivgi, and MiniMax together and Chinese labs account for about 54.8% of the video requests routed through OpenRouter that week. Runway's share is 0.3%.

Why the talking head wins

Request volume on a router measures automated work, not artistic ambition. And the single most automated video task in the economy right now is a person talking to camera about a product.

Think about what that job requires. A consistent face, a script, sync, a background, and a second language version by Thursday. It is generated in batches, one request per variant, and the variants are cheap because nothing in the frame moves except a mouth. A marketing team testing twelve hook variations for a paid campaign is twelve requests. A localisation job across six languages is six. Nobody watches the output twice, and nobody expects it to be beautiful.

Cinematic generation is the opposite shape of work. One shot, many attempts, high cost per attempt, human review between each. A studio making a 90 second sequence might run two hundred generations to keep eight. That is two hundred requests producing eight usable seconds, and it is a rounding error next to a brand shipping localized ad variants every week.

So the leaderboard is not saying HeyGen makes better video than Veo. It is saying more of the world's video work is templated talking heads than is short films, which anyone who has looked at a corporate content calendar already suspected.

The Chinese share is the real story

54.8% is a big number for a market where the loudest coverage is still American. ByteDance's growth is modest, up 6%. The interesting movement is Alibaba, up 152% week over week to 19.6%. That is what a price cut or a new integration looks like in this data: a step change, not a slope.

Wan 3.0 dropped 32% in requests in the same week its author share more than doubled, which sounds contradictory until you remember these tables cover slightly different windows and that mix shifts matter. A cheap fast tier can absorb volume that used to go to a flagship, and the request count falls while the share of spend rises.

MiniMax grew 23% on 5.2% share. SpaceXAI grew 32% on 11.4%. Google fell 17%. These are weekly swings on a router, and a single customer migrating a pipeline can move them. The pattern worth taking seriously is not any one week. It is that four Chinese labs now occupy four of the top eight slots, and that the cheapest fast tiers are where the volume is landing.

What Runway's 0.3% means

Runway at 0.3% is the number people will quote out of context. It does not mean Runway's models are unused. It means Runway's business is not routed through OpenRouter, because Runway sells subscriptions and enterprise contracts to people who use its own interface, and its API is not the default path for a developer who just wants a video call in their app.

That is a different go-to-market, and it is a legitimate one. It also means this leaderboard systematically undercounts anyone with a strong direct product. The same caveat applies to Google's 15.3%: Gemini and Flow users never appear in this table. There is a third distortion worth naming. A request is not a second of video. An avatar job might produce thirty seconds in one request. A cinematic job might produce four seconds per request and require fifteen of them to get a usable take. So a model that is expensive per attempt and used iteratively will look smaller here than its actual share of generated runtime, and a model that returns long finished clips in one call will look larger. The metric rewards products that bundle work into a single API call, which is exactly what an avatar product does.

Read the ranking as a map of developer defaults, not of total usage. For a builder, developer defaults are exactly the right thing to know, because your integration options, your community support, and your future deprecation risk all follow them.

What a buyer should do with this

If your job is templated talking-head content at volume, the data says the market has already picked, and the top two entries are avatar specialists rather than general video models. Test those first. They are optimized for a narrower task and the pricing reflects it.

If your job is a specific cinematic shot, none of this helps. Request volume says nothing about whether a model can hold a face through a camera move. You have to run your own footage, and you should expect to spend more than you budgeted on attempts.

If you are building a product that wraps a video model, the meaningful signal is tier proliferation. Seedance appears three times in the top ten at three different price points. That is a vendor telling you the base capability is settled and the competition has moved to cost per second. Build for a cheap tier and an upgrade path, not for one model string you fell in love with in July.

One more thing is worth testing before you commit: how the same prompt behaves across a vendor's two tiers. Seedance 2.0 Fast growing 89% while 2.0 stays flat suggests buyers are trading quality for throughput where they can, and that trade is only safe if the fast tier fails gracefully. Run a hard prompt through both. If the fast version does not just look cheaper but composes differently, your prompt library is not portable between tiers, and that is the kind of detail that only shows up in production.

And keep one habit from this data: the model at the top of the chart changed hands between weeks. Whatever you hardcode, put the model name in a config file, not in forty places across the codebase.