The Video Model Leaderboard Nobody Markets: Where the Requests Actually Go

Every week brings a new video generation ranking, and almost all of them are built the same way. Researchers gather outputs from competing models, put them in front of voters, and publish an Elo table. Those tables are useful for comparing quality. They say almost nothing about what buyers do.

A different kind of ranking is now public. OpenRouter publishes video model request volumes across its routing service, aggregated from real usage. The numbers for the week ending October 1 tell a story that the quality leaderboards do not.

The volume table

Google's Veo 3.1 Lite held the top spot with roughly 41,000 requests, up about 15 percent. ByteDance's Seedance 2.5 was second at about 30,000 requests, up around 28 percent. Seedance 2.0 Fast was third at 22,000. Veo 3.1 Fast was fourth at 17,000, up nearly 95 percent week over week. HeyGen Video entered the list as a new entrant with about 17,000 requests.

Below the top five, Alibaba's Wan 2.7 posted the largest change on the board, up more than 300 percent to about 12,000 requests. Wan 3.0 sat at roughly the same volume. Kling's Video v3.0 Pro entered the top ten at about 10,000 requests.

Now compare that to the quality tables. On Arena's text-to-video board, Gemini Omni 1.1 Flash leads. MiniMax H3 leads image-to-video. Kling 3.0 tops several commercial rankings on value. None of those models appears at the top of the request volume list.

Lite tiers are winning because the work changed

The pattern that explains most of the table is price and speed. Veo 3.1 Lite is Google's cheaper, faster tier, and it is the most requested model on the service. Veo 3.1 Fast nearly doubled its volume in a week. Seedance 2.0 Fast and Mini, both efficiency tiers, together account for more requests than the full Seedance 2.0.

That is what a market shifting from evaluation to production looks like. When people are testing a model, they pick the best one. When they are shipping something, they pick the one that makes the unit economics work. A model that produces a slightly less impressive clip for a fraction of the cost wins once the output is going into a pipeline rather than a demo reel.

The HeyGen Video entry is the same signal from a different angle. HeyGen launched its universal video model with a per-second price around a cent for the promotional period, roughly half the standard rate, and made it available through developer routers immediately. Within days it was in the top five by volume. Price and distribution, not benchmark position, produced that result.

Wan's jump is about openness

The largest week-over-week change belongs to Alibaba's Wan 2.7, which grew by more than 300 percent. Wan's family includes open-weight models, and open weights change how a model spreads. Anyone can host it, set their own price, and offer it as an option. When the weights are available, the model does not need the original vendor's permission to reach users.

That is the mechanism behind the open-weight video strategy. A closed model's volume is capped by the vendor's capacity and pricing decisions. An open model's volume is capped by how many providers choose to serve it. The second ceiling is usually higher.

Wan's growth also suggests something about where video generation is being used. Low-cost, self-hostable models tend to appear in high-volume applications: social content, e-commerce variations, game assets. Those do not need cinematic quality. They need enough output cheaply, and an open model served by many providers fits.

Usage and quality measure different things

None of this means the quality leaders are failing. It means they are competing for a different segment. A studio producing a commercial needs the best output and will pay for it. A platform generating thousands of clips a day needs cost per clip, latency, and availability.

The two segments are large, and they are pulling in opposite directions. That is why a single ranking of video models is misleading. A model can lead one table and be nearly invisible on the other, and both results can be correct.

There is a third consideration that volume data captures well: reliability. Request counts reflect completed work, and a model that produces a usable clip on the first attempt is worth more than one that requires three regenerations. Over millions of requests, that difference shows up as volume.

What the numbers cannot tell you

Request volume has blind spots, and it is worth naming them before drawing conclusions.

The data covers models served through one routing service, not the whole market. A lab that serves its own model to its own app, as several of the largest do, does not appear at all. A model that is popular inside a single large customer is undercounted. And request counts do not distinguish between one user generating a thousand clips and a thousand users generating one each, which is a meaningful difference for a vendor deciding where to invest.

The numbers also lag the launches. A model released last week and a model released last year can show the same volume while moving in opposite directions, and the percentage changes capture that only if you read them. Google's Veo 3.1 Fast nearly doubling in a week says more than its absolute rank does.

So the table is best read as a directional signal about where production traffic is going, not as a market share measurement. It answers the question of which models people are actually calling, which is a question the quality rankings never asked.

What the market looks like at the end of 2026

Four observations, drawn from the numbers rather than from any vendor's claims.

Fast tiers now drive the category. The most requested video models are the economical ones, which means the buyers with the most volume are optimizing on cost per usable second.

Open weights have a distribution advantage. Wan's growth rate is the clearest evidence that releasing weights expands the addressable market faster than a closed model can expand its own capacity.

New entrants can reach scale in days. HeyGen went from launch to a top-five position within a week by combining a low price, a developer-friendly integration, and an existing customer base. Distribution channels have become the fastest route to volume.

Quality rankings still matter for the high end. The models at the top of the Arena tables serve customers for whom a better clip is worth more than a cheaper one.

Anyone choosing a video model should read both. The quality tables say what a model can achieve, and the request volumes say what it costs to achieve it at scale. No single ranking captures both, which is one reason vendors keep publishing the quality tables.