Seven AI video models, one prompt: what we got for $8

Text-to-video models are now a pay-per-clip service: you send a sentence and get back a short film. But which one should you use, and what does a clip cost? We sent the same prompt to seven AI video models on the same evening and paid $8 for the lot. Two followed our direction closely. Three cut to a different shot halfway through. One delivered a higher resolution than we asked for, and one added sound we had switched off.

How we tested

We used OpenRouter's video API, which lets you reach many video models with one account and one kind of request: you submit a job, check on it until it is finished and then download the MP4. The owner of this site paid for the clips with his own OpenRouter credit.

Every model got exactly the same prompt:

We chose it because it combines things video models tend to find hard: hands, pouring liquid, a pattern that forms over time, rain on glass and a specific camera movement. We asked for 10 seconds in 16:9 at 720p, with sound. Where a model did not offer that, we took the closest option: Veo 3.1 Fast offers at most 8 seconds, and Hailuo 3 Max and HeyGen Video offer 768p. For Hailuo we switched sound off, because OpenRouter's list showed no sound for it.

The seven models were Seedance 2.5 (ByteDance), Kling 3.0 Pro (Kuaishou), Veo 3.1 Fast (Google), Grok Imagine Video 1.5 (xAI), Hailuo 3 Max (MiniMax), Wan 3.0 (Alibaba) and HeyGen Video, which HeyGen says is "built on MiniMax H3". Kling has already released Kling 4.0 according to its release notes; it was not on OpenRouter's list, so we tested 3.0. We left out Sora 2 Pro, the most expensive option on the list.

Each model got one try. Video models give a different result every time, so a second try with the same prompt can look quite different.

What each clip cost

OpenRouter returns the cost with each finished job. Together the seven clips cost $8.00:

Two prices were lower than OpenRouter's own list. HeyGen Video's list price at 768p is $0.03 per second, which would be $0.30; HeyGen's catalog page mentions a launch price of half the standard rate through October, which would explain the $0.15 we paid. Wan 3.0 is listed at $0.10 per second at 720p, so $1.00; we paid $0.85. Alibaba's price page mentions a limited-time discount, but we cannot say whether that explains this exact amount.

Waiting times are what our script measured from submitting to finished. It checked every 15 seconds, so the real times can be up to 15 seconds shorter.

Kling 3.0 and Seedance 2.5: one continuous shot, as asked

Two models did what the prompt described in a single, unbroken shot: the milk is poured, the latte art forms, and the camera pulls back until the rainy window and the bicycles come into view.

Kling 3.0 Pro comes closest to the prompt, as the filmstrip in this section shows. The leaf pattern forms step by step, the camera pulls back smoothly and ends on a barista by the window with a cyclist passing outside.

Seedance 2.5 also keeps one continuous shot, but its first second is odd: the milk is poured into black coffee without any foam, and the latte only appears after that. The pull-back to the window works, with bicycles visible in the background. It was also the most expensive clip in our test.

We judged this from the clips and from stills taken every second. We did not measure image quality in a technical sense, and with one try per model a different seed could change the picture.

Veo 3.1 Fast and Grok Imagine: smooth, but the latte was already done

Veo 3.1 Fast made an attractive, steady shot that moves from the cup to the window, where cyclists with umbrellas pass by, and it added croissants on the counter. But the latte art is already finished in the first frame, so the "forming a leaf" part of our prompt is missing. Veo offers at most 8 seconds, so this clip is two seconds shorter than the others.

Grok Imagine Video 1.5 also starts with finished latte art, while the milk keeps pouring for about six seconds while the pattern barely changes. Then the camera swings to the window and the bicycles. The movement is smooth, but it is a sideways move rather than the pull-back we asked for.

Both clips are usable as atmosphere shots. If the action itself matters, as in a product demo, they did less well with it than Kling and Seedance in our test.

Hailuo, Wan and HeyGen: cuts instead of a camera move

Three models solved "the camera slowly pulls back" by switching to a new shot. We measured this by comparing frames every 0.2 seconds: a sudden jump in the difference marks a cut.

• Hailuo 3 Max cuts after 4.6 seconds from the close-up to a wide shot of a table by the window. The pouring stops halfway through the pattern.
• Wan 3.0 fades from the close-up into a new shot at about 4.5 seconds. For a moment both images are visible at once, like a double exposure; then the cup stands alone on a counter in front of the window, with cyclists outside.
• HeyGen Video cuts three times: to a closer view of the cup at 4 seconds, to a barista behind the counter at about 8.4 seconds, and to cyclists outside at 9 seconds. It feels more like a short edit than a single shot.

None of this has to be a problem; a montage can be exactly what you want. But it is not what we asked for, and you have less control over the result. HeyGen was also by far the cheapest and fastest, which may make up for it if you need many short clips.

Not always what we asked for: resolution and sound

We also checked the files themselves. Two models delivered something other than what we requested:

• Kling 3.0 Pro returned a 1920 × 1080 video, although we asked for 720p. That fits the price we paid: $1.68 for 10 seconds matches the 1080p-with-sound rate of $0.168 per second on Kling's own price list. On OpenRouter this model is listed as 720p only, so if you choose it, expect Full HD and the matching price.
• Hailuo 3 Max delivered a clip with sound, although we had switched sound off. We measured the loudness of the audio track: it was clearly not silent.

All seven clips contained a stereo sound track. We did not judge the sound itself, such as whether the rain and café sounds fit the picture.

Who may use the result, and how it is labelled

Through OpenRouter, the terms say that "your ownership rights in the Output are set forth in the Model Terms for each Model you use". So the rules depend on the model. What we found on the makers' own pages on 4 October 2026:

• Veo (Google): Google's Veo documentation says videos are "watermarked using SynthID", an invisible mark. Google's Gemini API terms say it "won't claim ownership" of the content you create.
• Kling: the API terms for paid services say commercial use of generated content "is not restricted".
• Wan (Alibaba): the Model Studio terms say Alibaba does not claim ownership of the output, and ask users to mark it as AI-generated where the law requires that.
• Seedance: the API has an optional watermark that marks a video as AI-generated in the corner. We did not find terms on commercial use specific to Seedance.
• Hailuo (MiniMax): MiniMax's terms ask users to mark generated content clearly. We did not find a clear statement on commercial use.
• HeyGen and Grok Imagine: we did not check their terms for this article.

None of our seven clips showed a visible watermark. In the EU, rules on labelling AI-generated video are getting stricter, so if you publish such a clip, saying clearly that it is AI-generated is the safe choice. This summary is not legal advice; read the terms of the model you use.

How we checked

We generated the seven clips on 4 October 2026 through OpenRouter's video API, one try per model, with the prompt quoted above. Our script recorded every request, the status, the cost OpenRouter reported and the waiting time. We then opened the downloaded files in Chrome to measure length, size and sound, took one still per second, and compared frames every 0.2 seconds to find cuts. The descriptions of each clip are based on those stills and our viewing of the clips.

The videos in this article are the files exactly as the models returned them. The stills are frames from those videos, scaled down but not edited. Both are AI-generated and labelled as such. The charts and tables were made by us from our test records. The information on prices, rights and labelling comes from the makers' own pages as read on 4 October 2026: Kling, Google, Alibaba, ByteDance, MiniMax and HeyGen.

One try per model on one prompt is a snapshot, not a ranking. A different prompt or a second try could change the order, and prices and models change quickly.