What is available in the public beta of Modal Sandbox?

A managed Python image, a CPU sandbox with the lifetime and resource caps stated on this page, and GPU sandboxes across the tiers listed. Custom images and setup-time provisioning are not yet available.

How is the Modal Sandbox API priced?

One price to create a sandbox — flat for a short-lived CPU sandbox, higher by tier for a GPU or a longer lifetime — plus a small price per exec, status and terminate call. Lifetime is paid up front, and the 402 quotes the exact figure.

Can I get a GPU sandbox through the API?

Yes. Create takes a GPU type from the tiers on this page and prices the call by tier for the lifetime you request.

Does the sandbox keep state between commands?

Within one sandbox session, yes: create once, exec several commands, inspect stdout, stderr and exit codes between steps. Nothing persists after terminate.

Why run untrusted agent code in Modal rather than locally?

Model-written code is untrusted input. A sandbox keeps it off your host and your network, gives it a clean image every time, and is torn down explicitly so nothing is left running to be billed for.

Modal Sandbox via BlockRun vs E2B, Daytona or Modal direct?

All three sell isolated execution behind an account and a key. This is Modal's runtime reached per request with no account: an agent provisions a sandbox, runs its code and pays for that session alone.

How do I run agent code in a Modal sandbox without a Modal account?

Call BlockRun's sandbox endpoint. It creates a Modal sandbox, runs commands in it and tears it down, paid per request by the caller — no Modal account, key or plan.

Can an AI agent create its own sandbox unattended?

Yes. The create request pays for itself, so an agent with a funded balance provisions a runtime with no signup and no human in the loop.

Services By capability: Code Sandbox API

PARTNER

Modal Sandbox

Secure Code Runtime For AI Agents

Run untrusted Python in isolated sandboxes, keep multi-step workflows alive, and pay per run, priced before it starts. Modal gives agents a clean execution environment without making them manage containers, API keys, or infrastructure.

4 endpoints live on Base

Isolated execution

Pay per run

Persistent sandbox session

No account required

Website Modal Sandbox Docs

PUBLIC BETA

Public access today: Python 3.11, up to 1 vCPU, 1 GiB RAM, and a 5-minute sandbox lifetime. GPU sandboxes (T4 through H100) are live — see the pricing below. Custom images and setup-time provisioning are coming later.

Why Agents Reach For It

The value is not raw compute horsepower. It is giving an agent a safe, disposable, stateful place to run code when localhost execution would be risky or operationally messy.

Isolated Runtime

Run code outside your app server in an isolated sandbox session instead of on the host machine.

Stepwise Execution

Create once, execute multiple commands, inspect outputs between steps, and keep workflow state inside the sandbox.

Debuggable Workflows

Capture stdout, stderr, exit codes, and status checks so agents can reason about what happened and what to do next.

Wallet-Native Access

Pay per run with x402 on Base. No accounts, API keys, or infrastructure setup required for the caller.

What AI Agents Use It For

Modal Sandbox gives agents an execution layer they can call on demand, reason over, and tear down when the job is done — a safe, isolated place to run code without risking the host system.

Code Interpreter

$0.015 per session

Agent creates a sandbox, executes user-provided code safely, captures output, and terminates. Perfect for chat-based code assistants.

sandbox/createpython:3.11 image$0.011

sandbox/execrun user code$0.002

sandbox/terminatecleanup$0.002

Package Verification

$0.015 per test

Install dependencies in a clean sandbox, run checks or tests, and inspect failures without polluting the host environment.

sandbox/createmanaged Python 3.11 sandbox$0.011

sandbox/execpip install + pytest$0.002

sandbox/terminatecleanup$0.002

Data Processing & Automation

$0.015+ per job

Transform files, parse data, run scripts, or execute lightweight automation steps in a sandbox the agent can inspect and control.

sandbox/createup to 1 vCPU / 1 GiB / 5 min$0.011

sandbox/execrun processing step$0.002

sandbox/terminaterelease sandbox$0.002

Multi-Step Agent Workflows

$0.017+ per workflow

Create a persistent sandbox session, execute multiple commands across steps, inspect outputs between turns, and terminate when done.

sandbox/createcreate session$0.011

sandbox/execstep 1$0.002

sandbox/execstep 2$0.002

sandbox/terminatedone$0.002

Pricing

Two billing modes selected automatically by the requested timeout: short bursts pay a flat rate, long-lived sandboxes pay per hour.

Burst — timeout ≤ 300s

Flat rate per sandbox/create, covers the full short-lived session. Lifecycle ops at $0.002/request.

CPU sandbox/create

$0.011

per sandbox

Up to 1 vCPU, 1 GiB RAM, 300s lifetime

GPU sandbox/create

$0.05 – $0.4

per sandbox by GPU type

T4 $0.05 / L4 $0.08 / A10G $0.1 / A100 $0.2 / H100 $0.4. Up to 8 vCPU, 32 GiB RAM, 300s lifetime.

Long-lived — timeout > 300s, up to 24h

Per-hour rate × requested duration, billed upfront in one x402 settlement. No refund on early terminate — same as cloud-instance norms. Hours are exact (90min on T4 = 1.5 × $1.50 = $2.25).

Config	Per hour	Notes
CPU (1 vCPU / 1 GiB)	$0.10	Long-running batch jobs, scrapers, schedulers
T4 (up to 8 vCPU / 32 GiB)	$1.50	Light inference, small models
L4 (up to 8 vCPU / 32 GiB)	$2.00	Mid-tier inference
A10G (up to 8 vCPU / 32 GiB)	$2.50	Stable diffusion, 7B inference
A100 (up to 8 vCPU / 32 GiB)	$4.00	Training, 13B–70B inference
H100 (up to 8 vCPU / 32 GiB)	$8.00	Frontier training, low-latency 70B+

API Reference

Base (USDC): https://blockrun.ai/api/v1/modal/

All endpoints accept POST with a JSON body. Send without a payment header to get a 402 with the exact price and payment requirements.

Endpoint	Method	Price	Description
`/api/v1/modal/sandbox/create`	POST	$0.010	Create a managed Python 3.11 sandbox. Two billing modes selected by 'timeout': ≤300s = flat rate (CPU $0.01, GPU varies). >300s = per-hour billing for the full requested lifetime, no refund on early terminate. Max 24h.
`/api/v1/modal/sandbox/exec`	POST	$0.001	Execute a command inside a running sandbox. Returns stdout, stderr, and exit code.
`/api/v1/modal/sandbox/status`	POST	$0.001	Check the status of a sandbox (running or terminated).
`/api/v1/modal/sandbox/terminate`	POST	$0.001	Terminate a running sandbox and release its resources.

Quick Start

Create a sandbox session, run code, and clean up:

# Base (USDC on Base)
curl -X POST https://blockrun.ai/api/v1/modal/sandbox/create \
  -H "Content-Type: application/json" \
  -d '{"image": "python:3.11", "timeout": 300}'
# Returns: 402 with price ($0.011) and Base USDC payment instructions

# Paid request — create sandbox ($0.011)
curl -X POST https://blockrun.ai/api/v1/modal/sandbox/create \
  -H "Content-Type: application/json" \
  -H "x-payment: <x402_payment_token>" \
  -d '{"image": "python:3.11", "timeout": 300}'
# Returns: {"sandbox_id": "sb-xxx", "status": "running", ...}

# Execute code ($0.002)
curl -X POST https://blockrun.ai/api/v1/modal/sandbox/exec \
  -H "Content-Type: application/json" \
  -H "x-payment: <x402_payment_token>" \
  -d '{"sandbox_id": "sb-xxx", "command": ["python", "-c", "print(2+2)"]}'
# Returns: {"stdout": "4\n", "stderr": "", "returncode": 0}

# Terminate ($0.002)
curl -X POST https://blockrun.ai/api/v1/modal/sandbox/terminate \
  -H "Content-Type: application/json" \
  -H "x-payment: <x402_payment_token>" \
  -d '{"sandbox_id": "sb-xxx"}'
# Returns: {"sandbox_id": "sb-xxx", "status": "terminated"}

Use the BlockRun Python SDK or TypeScript SDK to handle x402 payment automatically.

Start running agent code safely

Fund your wallet with USDC on Base and give your agent a safe place to run code. No registration, no API keys required.

View Docs View All Services

Modal Sandbox API questions: isolation, GPUs, lifetime, pricing

What is available in the public beta of Modal Sandbox?: A managed Python image, a CPU sandbox with the lifetime and resource caps stated on this page, and GPU sandboxes across the tiers listed. Custom images and setup-time provisioning are not yet available.
How is the Modal Sandbox API priced?: One price to create a sandbox — flat for a short-lived CPU sandbox, higher by tier for a GPU or a longer lifetime — plus a small price per exec, status and terminate call. Lifetime is paid up front, and the 402 quotes the exact figure.
Can I get a GPU sandbox through the API?: Yes. Create takes a GPU type from the tiers on this page and prices the call by tier for the lifetime you request.
Does the sandbox keep state between commands?: Within one sandbox session, yes: create once, exec several commands, inspect stdout, stderr and exit codes between steps. Nothing persists after terminate.
Why run untrusted agent code in Modal rather than locally?: Model-written code is untrusted input. A sandbox keeps it off your host and your network, gives it a clean image every time, and is torn down explicitly so nothing is left running to be billed for.
Modal Sandbox via BlockRun vs E2B, Daytona or Modal direct?: All three sell isolated execution behind an account and a key. This is Modal's runtime reached per request with no account: an agent provisions a sandbox, runs its code and pays for that session alone.
How do I run agent code in a Modal sandbox without a Modal account?: Call BlockRun's sandbox endpoint. It creates a Modal sandbox, runs commands in it and tears it down, paid per request by the caller — no Modal account, key or plan.
Can an AI agent create its own sandbox unattended?: Yes. The create request pays for itself, so an agent with a funded balance provisions a runtime with no signup and no human in the loop.