Pruna is Pruna AI's model stack. The open-source library, the Pro token, and the API each print a different way to pay.
The library, then the Pro package
Pruna is the stack from Pruna AI GmbH. The repository is Apache-2.0. That install is the open-source package, pruna. The Pro guide says the paid package is a separate install, pip install pruna_pro==0.3.4, then you import smash from pruna_pro and pass a token. Pro adds advanced compression, an optimization agent, quality recovery, and priority support. A feature that belongs only to Pro is marked (Pro) in those docs. The guide says the core docs apply to both packages.
Buy Pro is not on the pricing page we read
The Pro guide tells you to open the pricing page and choose Buy Pro. Usage-based is US$0.40 an hour, as much as you need, billed at the end of the month. Credits are bought in multiples of US$200, and US$200 is printed as 500 hours of inference. Payment goes through Stripe. The token arrives by email. Questions go to [email protected].
The pricing page read for this listing does not print US$0.40 an hour, and it does not print Buy Pro. The rates on it are labeled Price from. Under the list it says inference providers can get preferential prices, and to make contact. It does not print that discount.
The API says there is no subscription
The API portal says no subscription is needed. You buy credits when you need them, and you can set a monthly usage limit. The homepage says the three ways to run Pruna are the API, self-hosted, or open-source. The open-source card says you deploy Pruna performance models on your own infrastructure, for control, compliance, or a custom setup.
One video name, three rates
Prices below are what each page prints. The homepage attaches a resolution. The pricing page says Price from and does not name the resolution. The model guide prints a range.
- p-video-2-pro: the pricing page says Price from US$0.01 a second. The homepage says US$0.02 a second at 480p. The model guide says US$0.02 to US$0.075 a second, by resolution and mode, and 50% off until 13 October 2026. It also says this model is built on Minimax H3. The homepage and the pricing page do not print that date.
- p-video: the homepage at 720p and the pricing page both say US$0.02 a second. The model guide says US$0.005 to US$0.04 a second, by resolution and draft mode.
- p-video-2: the homepage at 720p and the pricing page both say US$0.025 a second. The model guide says US$0.015 to US$0.05 a second, by resolution and draft mode.
- p-video-edit: the homepage at 720p and the pricing page both say US$0.045 a second. The model guide says US$0.025 to US$0.045 a second.
The same model guide also bills some videos by the clip, not by the second. vace is US$0.40 a video. wan-t2v is US$0.05 a video at 480p and US$0.10 a video at 720p. wan-i2v is US$0.05 a video at 480p and US$0.11 a video at 720p.
Three more video names split by resolution. The homepage prices p-video-avatar, p-video-animate, and p-video-replace at 720p: US$0.025, US$0.03, and US$0.03 a second. The model guide repeats those 720p rates and adds 1080p at US$0.045, US$0.06, and US$0.06. The pricing page's Price from lines for those three are US$0.025, US$0.03, and US$0.03 a second, and they do not name a resolution.
Still images line up more closely. The model guide prices p-image and p-image-lora at US$0.005 an image, and p-image-edit and p-image-edit-lora at US$0.010. The pricing page prints US$0.005 and US$0.010 per image output, and one of the US$0.005 lines spells output as ouput. The two trainers are capped at 5 requests a minute in the model guide.
Garments, megapixels, and a 30-minute file
Try-on is printed in two units. The pricing page says US$0.015 for the first garment and US$0.008 for each additional garment. The model guide says 15 units for the first image input and 8 units for each additional one. Those lines are left as printed.
Upscale on the pricing page is US$0.10 per 8MP. The model guide says US$0.005 to US$0.12 an image, by the target megapixel count. Training agrees on both pages: US$1.80 per 1,000 steps, and US$4.00 per 1,000 steps. The model guide adds that those training jobs are async only, and that the output expires in about 30 minutes.