Pricing
Open weights on our own cards, from $0.003 an image and $0.018 for five seconds of video. Billed per second, per image or per million characters, with no plan and no seat.
Video
Billed by the second of video, at the resolution you ask for.2 modelsLTX-2.5
LightricksLightricks' open weight world model, which generates synchronized video and audio from text or from an image. We run the distilled transformer in full bf16 precision with its Gemma 4 12B text encoder on RTX PRO 6000s, on the fixed 8 step schedule Lightricks ships. Hugging Face keeps the weights behind a contact agreement, so running them yourself means accepting their licence first.
- 480p$0.004fal —
- 720p$0.007fal —
- 1080p$0.018fal $0.06
- 1440pnot yetfal $0.12
- 4Knot yetfal $0.24
A five second clip at 1080p costs $0.09 here, $0.30 at fal.
- Limits
- Up to 8 s per render, 24 fps
- Runs on
- RTX PRO 6000
Wan 2.2
Alibaba Wan-AIAlibaba's Wan 2.2 in its five billion parameter size, Apache-2.0, run from the diffusers build its authors publish. It renders picture only: there is no audio track, where our LTX tier renders sound with the image.
- 1280 x 704$0.049Novita $0.10fal $0.05Replicate $0.25
A five second clip at 1280 x 704 costs $0.245 here, $0.50 at Novita.
- License
- apache-2.0
- Limits
- 1280 x 704, 24 fps, 4 to 8 s per render, no audio
- Runs on
- H100
Image
Billed by the image, at the size you ask for.2 modelsFLUX.2 klein
Black Forest LabsThe four billion parameter FLUX.2, the size Black Forest Labs publishes under Apache-2.0. Their larger builds carry licences we do not accept, so this is the one we run.
- Standard, up to 1024 px$0.003DeepInfra $0.014Replicate $0.025
- Large, up to 1536 px$0.006DeepInfra $0.032Replicate $0.025
- Max, up to 2048 px$0.012DeepInfra $0.056Replicate $0.025
Ten images at 2048 px cost $0.12 here, $0.56 at DeepInfra.

- License
- apache-2.0
- Limits
- Up to 2048 on the long side, sides divisible by 64
- Runs on
- H100
Qwen-Image
Alibaba QwenAlibaba's Apache-2.0 image model, and the one in this catalogue that writes readable text inside a picture. It is also the heaviest, which is what its price says.

- License
- apache-2.0
- Limits
- Up to 2048 on the long side, sides divisible by 64
- Runs on
- H100
Speech
Billed by the million characters of script.1 modelVoxCPM2
OpenBMBOpenBMB's tokenizer free, diffusion autoregressive text to speech model, Apache-2.0 and free for commercial use. Its model card states 2B parameters, 30 languages and 48 kHz output. We run it on a single L40S and charge the card time a narration actually takes.
- License
- apache-2.0
- Limits
- Up to 40,000 characters per request, 48 kHz output
- Runs on
- L40S
Credits
You load a balance and spend it, render by render. There is no plan, no seat and nothing that renews on its own.
- $10555 clips of 5 s, or 3333 images
- $251388 clips of 5 s, or 8333 images
- $502777 clips of 5 s, or 16666 images
Any amount between $5 and $500 works. We run open weights on on-demand cards and take a small margin over the card time, which is the whole reason these prices are what they are.
Included, at no extra cost
- No subscription
- You load a balance and spend it. Nothing renews.
- The price you see
- What the page says is what leaves your balance.
- Refused, refunded
- If the engine refuses a render, the charge comes back.
- Open weights
- Every licence is named and linked, on every model.
- One queue
- There is no paid fast lane. Nobody passes in front.
- Render from Claude
- The MCP endpoint bills the same balance.
Questions
- What happens if a render fails?
- If the engine refuses the request, the amount is credited back to your balance the same second. Every movement is listed under Activity.
- Does the resolution change the price?
- Yes. A second of 1080p holds the card about two and a half times longer than a second of 480p, and the rate follows. Every step is in the table above, before you run anything.
- Is there VAT on top?
- VAT not applicable, article 293 B of the French General Tax Code
- Is there a minimum per render?
- A small one. Loading the weights costs the same whether the clip runs one second or eight, so a render never bills below that floor. It is already in the table: the shortest clip is where the rate and the floor meet.
- Do credits expire?
- Your balance stays until you spend it. We do not clear it on a timer.
- Why is the first render after a pause slower?
- A card that has been idle has to load the weights again, which adds a few tens of seconds. It costs you nothing extra: you pay per second of video, not per second of card.
- Can I switch models without changing my code?
- Yes. One endpoint, one body, and the model is a field. Swap the slug and nothing else moves.