Vercel Sandbox stops billing internet downloads for agent runs

Vercel has removed one of the easiest ways for agent sandboxes to produce surprise bandwidth usage: data downloaded from the internet into Vercel Sandbox is now free. The change covers package installs, Git repository clones, artifacts, and datasets pulled into a sandbox during a run.

The limit matters because sandboxed agents often spend their first minutes fetching context before they do useful work. A coding agent might install dependencies, clone a monorepo, pull fixtures, download model test data, and fetch build artifacts before it writes a line of code. Under the new rule, those inbound downloads no longer count toward Sandbox Data Transfer usage.

Vercel removes inbound download charges from Sandbox

Vercel's changelog says Sandbox no longer bills for data it downloads from the internet. The official pricing documentation now makes the boundary explicit: data a sandbox downloads, including packages, Git repositories, artifacts, and datasets, is free.

That is a pricing change, not a full free-for-all. Traffic received on exposed ports remains billable, and data the sandbox sends out to the internet still counts. Active CPU, provisioned memory, snapshot storage, and sandbox creation pricing are unchanged.

For agent builders, that distinction is practical. A dependency installation, git clone, or dataset pull should be cheaper to run repeatedly. A sandbox that serves a preview app to external users or pushes large files to another service still needs bandwidth controls.

What agent and CI teams should reprice

The affected workflows are the ones that load external material at runtime. Common examples include:

• AI code validation that installs package trees before tests.
• Repository analysis agents that clone fresh project state for each task.
• Evaluation jobs that pull public datasets or artifacts into disposable environments.
• Build sandboxes that fetch large dependency caches from the public internet.

Vercel's pricing page still lists five metered dimensions for Sandbox: Active CPU, Provisioned Memory, Sandbox Creations, Sandbox Data Transfer, and Snapshot Storage. The change narrows what Data Transfer means. Incoming downloads from the internet are excluded, while outbound traffic and exposed-port traffic remain in the meter.

The official examples also show that CPU and memory can dominate small runs. A short validation job may cost only cents on Pro or Enterprise, but long-running agents still accumulate provisioned-memory and active-CPU usage. Teams should recalculate based on the whole run, not only on bandwidth.

The billing line is useful, but still sharp

The new line helps with a common agent pattern: disposable environments that fetch a lot and return a small result. It does less for sandboxes that expose services, proxy traffic, run long previews, or upload large generated artifacts.

Hobby users still have included monthly usage and sandbox creation pauses after the included limit is exceeded. Pro usage runs against the monthly credit before overage rates apply. Enterprise limits can vary by contract.

The safest migration step is boring and useful: separate inbound setup traffic from served or uploaded traffic in your cost dashboards. If a workflow mostly downloads dependencies and produces logs or patches, this Vercel change may cut one cost source. If it streams previews or exports large assets, it still needs explicit bandwidth limits.

Sources and methodology

This post uses Vercel's July 17 changelog as the primary source, Vercel's Sandbox pricing documentation for the billing boundary, and DevBytes as independent corroboration that the change is visible outside Vercel's own site. DevBytes describes the same core split: internet downloads are free, while incoming exposed-port traffic and outbound internet traffic remain billable.

For teams building agent workflows, the immediate action is to update cost assumptions for setup-heavy sandbox runs and keep outbound or exposed-port traffic under review.