Pika H3 Flex: what it is, pricing and how to use it
2026-10-31 · 4 min read
Pika H3 Flex is a new video model from Pika that makes clips of 5 to 15 seconds with synced stereo sound, starting from a text prompt, a first or last frame, or up to 12 reference files. Pika's hosted version costs $0.01 to $0.16 per output second at launch prices, depending on resolution. The weights are also posted on Hugging Face, but read the license before you run them in the United States.
Pika launched it on October 29. The @pika_labs announcement says it debuted "as the #1 open weight model on DesignArena's t2v board," and claims a 15-second clip with a soundtrack can render in as little as 7 seconds. A follow-up post linked the GitHub code, the weights and a ComfyUI tutorial on day one. Co-founder @demi_guo_ summed up the pitch as "More creative controls, leading quality, and cheapest cost!"
What Pika H3 Flex is
Pika didn't train this model from scratch. It took MiniMax's H3 video model and post-trained it. In a technical thread on October 30, Pika says it cut generation from 50 denoising steps to six and then tuned the result for quality and control. Fewer steps is where the speed comes from. Pika's FAQ adds that the 7-second number measures model compute alone on a 15-second 768p clip, so end-to-end API time will run longer.
On DesignArena's text-to-video board, Pika's launch page puts Flex at #7 overall, ahead of base MiniMax H3, H3 Max and Seedance 2.5. That ranking is Pika's launch-day reading, and arena boards move around a lot.
What it can do
The Hugging Face model card and Pika's product page list these modes:
- Text to video with audio
- First frame, last frame, or first and last frame to video
- Reference to video using up to 9 images, 3 videos and 3 audio clips (12 files total)
- Clips of 5 to 15 seconds at 24 fps, in 11 aspect ratios from 21:9 to 9:21
- Stereo 32 kHz sound mixed into the MP4
Some features only run on Pika's servers. You get 1080p, 2K and 4K upscaling only through the API, and the same goes for splitting the soundtrack into five WAV stems (dialogue, score, SFX, foley and ambience). I'd argue the stems are the most practical feature here, since you can swap the music without regenerating the clip. If you need stand-alone sound design, our ElevenLabs SFX v3 guide covers another option.
Pika H3 Flex pricing
Pika charges per second of output video. These prices are from its product page:
| Resolution | Regular price per second | Launch price per second |
|---|---|---|
| 480p | $0.02 | $0.01 |
| 768p | $0.04 | $0.02 |
| 1080p | $0.07 | $0.035 |
| 2K | $0.08 | $0.04 |
| 4K | $0.16 | $0.08 |
At launch prices, a 15-second 768p clip comes to about $0.30, and the same clip in 4K is about $1.20. Reference-to-video costs more and bills on both input and output video seconds ($0.03 per second at 480p during launch). Image and audio references are free. Stems add $0.001 per second.
Pika's page also offers API Club members 15 free Flex generations a day from October 29 to November 5. According to the press release, the model reaches the regular pika.art creation app on November 2.
How to use it
The quickest route is the Pika API Club at dev.pika.art. Make an account, use some of your free daily generations to compare a text prompt against a first-frame prompt, and then try a reference job with a product photo and a short voice clip.
To self-host, start with the GitHub repo. It includes diffusers and ComfyUI workflows. Be ready for the hardware. The release ships as LoRA adapters on top of roughly 210 GB of MiniMax H3 base files, and the model card says Pika's test requests passed on an H100 with 80 GB. Local runs top out at 480p and 768p.
The license catch
Pika calls the release open source, but the adapters ship under the MiniMax H3 Community License, and that license lists "the European Union, the United Kingdom, the Republic of Korea and the United States of America" as excluded territories. It also bars using the works or their outputs outside the permitted territory. Read plainly, a US business can download the weights but can't run them in-house under that license. The hosted API is a separate commercial product with its own terms, so for American shops it is the realistic way in. If you want to see what running a big open model locally involves, our Mistral Large 4 open weights guide walks through it.
What it means for a small business
Short product clips, social ads and explainer loops can now cost a few cents to a dollar each. At that price you can afford a dozen drafts before you pick one, so the generation step stops being the bottleneck. The slow part becomes deciding what to make, who signs off and where it gets posted. If you want help mapping that, New Face Design's free process audit is built for it.