Step-by-step guides
5 min read4 decisionsHow much CPU, RAM and disk an AI sandbox really needs
A Sandbox is a powerful computer in the cloud: you switch it on when you need it, it pauses itself when you do not, and your files stay on its disk. Creating one takes four decisions, and none of them is final.
GPU Flow team. Updated on 9 November 2026.
This guide is about getting it right the first time without overpaying. Detailed prices are on the pricing page; here is the reasoning.
The GPU is not inside, and that is good news
It is the most expensive question and it has a short answer: a Sandbox has no graphics card inside. If your work is an app or an agent that calls a language model, it does not need one: the model runs on the cluster GPUs through Token Factory, and your Sandbox reaches them over the internal network, never hitting the public internet.
So you pay for the computer by the hour and for the models by what you use, instead of renting an expensive card that sits idle while you type. If you genuinely need the GPU inside (training your own model, accelerated video processing), that is dedicated GPU today, a different product of ours.
Pick the size
There are three, and they fix the processors (vCPU) and memory (RAM) the machine gets. Start small: if it gets tight, create a bigger one with the same disk and you lose none of your work.
| size | processors | memory | running | for |
|---|---|---|---|---|
| Starter | 4 vCPU | 8 GB | 0.15 €/h | Trying things out and small projects |
| Pro | 8 vCPU | 32 GB | 0.30 €/h | Day-to-day work, with room to spare |
| Team | 16 vCPU | 64 GB | 0.55 €/h | A team sharing the same environment |
Pick the disk
The disk is where your files live and it does not depend on the size you pick: you choose it, between 50, 100, 250, 500, 1,000 GB, starting at 50. Size it by your repository, its dependencies and the data you work with.
It is the only thing still billed while the Sandbox is paused, because it is what keeps your things where you left them. You can grow it later, so you do not need to guess today what you will need in a year.
Pick an assistant, if you want one
You can boot the Sandbox with a coding assistant already installed and connected to Token Factory, with no keys to configure by hand. With Claude Code you choose between using your own Anthropic account or paying per token with GLM 5.3 NVFP4, the fixed model for that template. With Open Code you pick the model from inside the environment and pay for the tokens it uses.
If you would rather set it up yourself, you can: it is a normal Ubuntu machine you reach over SSH.
Auto-pause is what saves you
When you stop using it, the Sandbox pauses itself and stops counting the hourly rate, so keeping one around all week does not ruin you: you pay the hours it was running. When you come back your files are still on its disk; whatever was in memory is not. The disk is always billed, running or paused, and it is your minimum spend for the month.
Create your Sandbox
Pick size, disk and assistant, and it boots in a minute. The hours are charged to your card or, if you choose so in Settings, to your Token Factory credit.
Related
- Run Claude Code with open-source models (Qwen, GLM): setup with environment variablesPoint Claude Code at Token Factory with environment variables, or boot it ready-made in a Sandbox, with the open-source models in the catalogue.
- GPU Flow vs Modal: GPUs, sandboxes and inferenceWhat a B200, a sandbox for an agent and inference cost on each, with Modal’s prices checked on its own site.