Simple, prepaid pricing.
Add credit by card and pay only for what you use. No subscription, no minimum term, and credit does not expire.
What one GPU hour makes
Our ready programs run on a single GPU machine. One hour of that machine, used non-stop by you or your AI agent, makes:
| Program | Output (default settings) | Per GPU hour | Machine price | Works out to |
|---|---|---|---|---|
| FLUX 3 Klein | Images832×1216 | about 500 images | about $1.00/hour | $0.002per image |
| Krea 2 | Images832×1216 | about 400 images | about $1.05/hour | $0.003per image |
| LTX-2.5 | Video with sound544×960, 24 fps | about 12 minutes of video | about $2.00/hour | $0.17per minute of video |
| MiniMax H3draft | Video with sound352×608, 24 fps | about 13.5 minutes of video | about $2.00/hour | $0.14per minute of video |
| MiniMax H3full quality | Video with sound704×1216, 24 fps | about 4.5 minutes of video | about $2.00/hour | $0.44per minute of video |
You control size, length, shape and quality in every request; bigger and longer outputs take more seconds, and you pay for exactly those seconds. Machine prices follow the GPU market and are shown before you start.
How this compares
Most AI services charge per image or per second of video, at a fixed list price. When you know what you want to make, running the model on your own GPU hour is a different order of cost:
| What you make | Public pay-per-output price | On GPU Forge |
|---|---|---|
| A 10-second video with sound | about $4.00 (Google Veo 3.1 Standard, about $0.40 per second) | the same $4.00 buys about 28 minutes of MiniMax H3 draft video or about 23 minutes of LTX-2.5 video |
| One minute of video with sound | about $7–24 (Google Veo 3.1, Fast to Standard) | $0.14–0.44 |
| One image | about $0.02 (FLUX Krea on typical per-image APIs) | about $0.003 |
Public list prices as of November 2026. Models, resolutions and quality differ between services: these are different models, and Veo renders at higher resolutions. The point is the price model: with GPU Forge you pay for GPU time, not per output, so large and repeated work costs a small fraction.
Two ways to pay, the same low rate
Pay per result (jobs)
Send one request, get one image or clip. You pay only the seconds of rendering it used, at the machine's hourly rate, with a quote before it runs and an optional price cap (minimum $0.002 per job). A job that fails is not charged, and the wait while a machine wakes is never billed. No machines to manage.
Your own machine
Rent the program's machine, billed per second, and send it as much work as you like: the more you generate in an hour, the less each result costs. Stop it and it parks with its disk kept for a small storage price, ready again in about 20–60 seconds.
GPU machines for your own software
Billed per second at the hourly price shown before you start, across Vast.ai, RunPod, Nebius and Jarvis Labs. For example:
- RTX 5090from about $0.55an hour
- B200from about $8an hour
Your balance
When your balance reaches $0, nothing new starts, and machines still running are released about 10 minutes later, disks included. Automatic top-up and daily spending limits keep this from happening by surprise.
Refunds
Unused balance is refunded on request to your original card, usually within 10 business days. Write to support@gpuforge.io.
Start with prepaid credit
Sign in with Google or GitHub and add credit by card. No subscription, no minimum term, and credit does not expire.