Sandboxes
Every agent runs in its own isolated sandbox. You pick the sandbox provider and size when you create the agent, under Sandbox Provider and Sandbox Size. The provider is fixed once the agent exists.
Providers
| Provider | Sizes | Shared folders | Notes |
|---|---|---|---|
| Daytona | S, M, L | Yes | The default. Offers GPU sandboxes. |
| E2B | L | No | |
| Cloudflare | S, M, L | Yes | Sleeps when idle and resumes with its workspace intact. Coding agents only. |
| Modal | L | No | Experimental. Sandboxes live 24 hours at most, and a task does not resume afterwards. Coding agents only. |
Which providers appear in the picker depends on your workspace. See Sandboxes in the API reference for every provider the API accepts and which runtimes each one runs.
Pricing
Sandbox time is paid from your Brainbase credits, the same balance that pays for model usage. One credit is $0.01. Prices are per hour of active sandbox time.
| Provider | S | M | L |
|---|---|---|---|
| Daytona | $0.084 / hr | $0.167 / hr | $0.332 / hr |
| Cloudflare | $0.129 / hr | $0.220 / hr | $0.401 / hr |
| E2B | — | — | $0.60 / hr |
| Modal | — | — | $0.60 / hr |
Prices here are estimates and may change.
- Sizes: S is 1 vCPU and 2 GiB of memory, M is 2 vCPU and 4 GiB, L is 4 vCPU and 8 GiB. Each has 10 GiB of disk.
- Daytona is billed per second for the CPU, memory, and disk the sandbox is given.
- Cloudflare is billed per second for the Cloudflare instance that hosts the size, which rounds memory and disk up: S runs on 1 vCPU / 6 GiB / 12 GB, M on 2 vCPU / 8 GiB / 16 GB, and L on 4 vCPU / 12 GiB / 20 GB.
- E2B and Modal are a flat 1 credit ($0.01) per active minute.
GPU sandboxes
GPU sandboxes run on Daytona at size L with one GPU. You cannot choose the GPU type. The GPU is billed per second, on top of the sandbox's CPU, memory, and disk, at the rate for the type the sandbox gets.
| GPU type | Price per GPU |
|---|---|
| NVIDIA H100 | $3.95 / hr |
| NVIDIA H200 | $4.54 / hr |
| NVIDIA RTX PRO 6000 | $3.03 / hr |
| NVIDIA RTX 5090 | $1.29 / hr |
| NVIDIA RTX 4090 | $0.99 / hr |
How billing works
- Active time only: The meter runs while the sandbox is active. A sandbox that is stopped or asleep is not charged for that time.
- Separate from model usage: Sandbox time is charged in addition to the model calls the agent makes. With Bring Your Own Bedrock, model calls go to your AWS account but sandbox time still uses credits.
- Out of credits: A team with no credits left cannot run tasks until credits are added.