What a minute of AI video costs in Beijing vs. San Francisco
Real per-second prices for Veo 2, Kling 2.0 and Jimeng this month, and why the retake rate matters more than the sticker price.
People keep telling me AI video is “getting cheap” or “still too expensive,” and almost nobody gives numbers. So I sat down with the public prices this week. I’m using public list prices, without enterprise discounts or information from my job.
I’ll use 5-second clips and per-second prices as the unit, at roughly 7.3 RMB to the dollar.
Google Veo 2, via the API, was priced in February at $0.50 per second. At that rate, generating 60 seconds of footage costs $30 before any retakes.
Kling 2.0, released by Kuaishou on the 15th, is the best-looking model I’ve tried this month. On the Chinese consumer plan (Gold membership, about 46 RMB a month for 660 credits), a 5-second Kling 2.0 clip costs 100 credits. So the plan buys you six and a bit clips a month, and each one works out to around 7 RMB, which is about 1.4 RMB, or roughly $0.19, per second. The older standard model on the same plan costs a tenth of that.
Jimeng, ByteDance’s creative app, is much cheaper. Its basic membership, about 69 RMB a month for 1,080 credits, generates on the order of 360 short videos a month. That’s well under 0.1 RMB per second, a cent or less per second in dollars. The quality is below Kling 2.0 for the best shots, but for a lot of ordinary content it’s usable.
So the raw cost of one minute of generated video ranges from about $30 with Veo 2, to about $11 with Kling 2.0, to well under a dollar with Jimeng. That’s a factor of more than thirty between the most and least expensive, for tools that all produce something you’d call professional-looking.
But the sticker price is misleading, because you never use every clip. The real unit is the cost per usable second. For anything with characters, you retake a lot: a hand goes wrong, a face drifts, the action doesn’t match the prompt, the camera does something you didn’t ask for. In my own tests with narrative scenes, I’ve needed anywhere from three to ten generations to get one clip I’d keep. Multiply accordingly. A cheaper model with a worse keep rate can end up costing more per usable second than an expensive one that gets it right the second time.
Price competition in China is intense and it’s pushing the floor down fast. There are many companies with competitive video models there, several treating generation as a creator acquisition tool rather than a profit center. I’d expect Chinese consumer prices to stay far below American ones for a while.
The keep rate is the metric that matters, and nobody publishes it. Better consistency and control improve economics more than a lower per-second price does.
For a short video, generation may already be a smaller expense than editing. Even at Veo’s rate, the raw footage for a minute costs $30 before retakes; choosing and assembling usable shots still takes time. The constraint is whether the output is good enough, coherent enough across shots, and controllable enough that someone with taste can make something worth watching.