AI Potluck
Infrastructure / Deployment

Modal Sandboxes

Modal Labs

Modal Sandboxes runs untrusted or agent-generated code in gVisor-isolated containers on Modal's serverless cloud, created at runtime from the Python, JavaScript or Go SDK. A sandbox can be built from an arbitrary container image, looked up later by name, given GPUs, and held open for anything from the five-minute default to twenty-four hours. Compute is billed by the core-second.

Modal bills Sandbox compute at a separate rate from its Function compute, so the plan tiers on the pricing page describe the surrounding platform rather than the sandbox itself. Verified 2026-08-13 via the Modal Sandboxes guide and the Modal pricing page.

Openness

1 high confidence
1.0
license
Proprietary
source
closed
service
proprietary Modal managed cloud (no self-host of the gVisor execution plane)
open-part
client SDK only (modal Python/JS client libraries)
pricing
pay-as-you-go usage, free starter credits

Closed managed service. The client SDK is open, but the execution substrate is a proprietary managed cloud with no self-host option, and an open client over a closed runtime is not open core. The contrast with E2B is the point: E2B ships self-hostable infrastructure, Modal does not.

  • https://modal.com/pricing recorded 2026-06-04

    Modal is a managed proprietary cloud platform, pay-per-CPU-cycle, $30/mo free credits

  • https://modal.com/docs/guide/sandbox recorded 2026-06-04

    Sandboxes = secure containers for untrusted agent code, driven via Modal SDK

  • https://modal.com/pricing recorded 2026-08-13

    Three plans - Starter ($0 + compute, $30/mo free credits), Team ($250 + compute) and Enterprise (custom) - all metered against Modal's own cloud, with a separate Sandbox and Notebooks compute rate per core-second and GiB-second. Nothing on the page offers a self-hosted, on-premises or source-available deployment.

  • https://modal.com/docs/guide/sandbox recorded 2026-08-13

    Redirects to /docs/guide/sandboxes. Sandboxes are still "secure containers for executing untrusted user or agent code on Modal", created only through the Python, JavaScript or Go SDK against Modal's service; no runtime is distributed to run elsewhere.

Adoption

3 low confidence
3.0

No Modal Sandboxes user or usage count is disclosed on the primary pages; Modal is a well-known managed AI-compute platform with broad developer use and a documented free tier. Placed at 3 on reported multi-customer traction rather than on a measured figure, kept conservative in the absence of a hard number. The nearest figure Modal publishes is platform-wide rather than per-product - its sandbox comparison page claims Modal "powers infrastructure for over 10,000 teams" and names Lovable and Quora as production users running untrusted code - which is another reported-traction claim rather than a count of Sandboxes use, so no reach figure is recorded.

Capability

4 medium confidence
4.0

Strong, purpose-built agent sandbox: gVisor isolation, arbitrary images and high concurrency. The Sandboxes guide documents arbitrary custom images, named lookup and a 5-minute default lifetime extendable to 24 hours, and Modal’s own comparison page states gVisor isolation and 50,000+ concurrent sessions. It sits just below the Firecracker-microVM frontier (Lambda, Vercel) on isolation strength, so 4 - level with the Ray and Ollama capability anchors. Independent ComputeSDK TTI benchmarks put Modal at a 0.76s median cold-start, 0.82s P95, composite 92.1, 14th by median of 24 providers, against ~470ms and 13th of 19 on the earlier run: slower in absolute terms and unchanged in placement, still mid-field, still short of the sub-0.3s leaders and well clear of the multi-second tail.

Verified 2026-08-13