Pika has retrained MiniMax's H3 video model so that it builds a clip in six passes instead of 50. The result is a model that makes a 15-second draft with sound for about $0.60, or half that while launch prices last. It is available now through Pika's API and comes to Pika's own app on November 2.
Making video stops being the expensive step
A video model starts from noise and cleans it up over many passes, and every pass costs computing time. Going from 50 passes to six is where the lower price comes from. At 768p the regular price is $0.04 a second, so twenty versions of a 15-second ad cost $12. Finishing the best one in 4K is billed as the difference between the two prices, which is about $2 more.
That is the price of a lunch for twenty options and a finished file. A school department that wants a short welcome video, or a recruiter who wants a clip about life at the company, can now compare many versions before committing to one. Email already works this way: nobody pays per draft, so people write three and keep the best. Video was too expensive to be treated like that, and for short clips it no longer is.
What was already there, and what Pika added
Most of the headline features come from the model underneath. MiniMax's H3 was already open weights and took images, video and audio as references. Flex adds speed and price, a separate finishing pass, and sound you can split apart.
The open part has limits. The weights for the 480p and 768p generator are on Hugging Face, so a company with enough hardware can run it privately, which matters if its unreleased product photos cannot go to someone else's service. The 4K finish and the audio stems run only on Pika's service.
Sound you can fix without remaking the picture
A generated clip normally comes with one mixed soundtrack. If the music is too loud or the footsteps are wrong, the choice is to live with it or generate the whole clip again. Through Pika's API you can ask for dialogue, score, sound effects, foley and background ambience as five separate tracks. A manager making a training video can turn the music down under the voice and leave everything else alone, in an ordinary editing program.
What does not get cheaper
Pika's seven seconds for a 15-second clip is the model's own computing time. Waiting in the real service takes longer, especially with references or a 4K finish. A video reference is also billed for its own seconds on top of the new clip, so a long motion reference raises the price of every attempt.
The larger cost moves to the person looking. Twenty drafts mean twenty things to watch, and a model that is happy to make anything will make plenty that are wrong for your brand. The reference files help: a product photo keeps the item looking like the real one, and you tell the model what each file is for. But someone still has to know what a good version looks like for the people who will see it. When drafts cost almost nothing, that judgment is what the work is made of.