Modal vs Vercel Sandbox: pricing, isolation and limits

Modal isolates with gVisor and bills requested CPU; Vercel uses Firecracker and bills active CPU, which is cheaper for agents that wait.

Against these two, Runtime cuts the bill by 84% and 70%. One job, 1,000 one-minute runs of a 2 vCPU, 4 GiB sandbox, costs $3.97 on Modal, $2.12 on Vercel and $0.64 on Runtime. Modal charges for the CPU a sandbox reserves; Vercel and Runtime charge for what it uses, and Runtime's rate for that is about a fifth of Vercel's. Runtime also gives each sandbox a Firecracker microVM kernel of its own. Rates checked 23 September 2026.

At a glance

Vercel's figures are at the rates of its default iad1 region.

	Modal Sandboxes	Vercel Sandbox	Runtime
Isolation	gVisor, a user-space kernel that intercepts system calls	Firecracker microVM, own kernel	Firecracker microVM, own kernel
CPU	$0.1419 per physical core-hour (2 vCPUs), request or use	$0.128 per active CPU-hour	$0.025 per vCPU-hour of measured CPU, with a small floor
Memory	$0.0240 per GiB-hour, request or use	$0.0212 per provisioned GB-hour	$0.0075 per reserved GiB-hour
Plan fee	Starter $0 with $30 a month of compute; Team $250 a month	Hobby allowance free; usage beyond it needs Pro	None; prepaid credit from $10
Free start	$30 of compute a month on Starter	5 active CPU-hours and 420 GB-hours a month on Hobby	50 sandbox hours, no card
Session length	Up to 24 hours; snapshots carry state beyond that	45 minutes on Hobby, 24 hours on Pro	Leases of up to an hour, extended as often as needed
Keeping state	Snapshots	Stop keeps the filesystem; processes start again	Pause keeps files, memory and processes, 1 to 365 days
Setup	A Modal App is needed outside Modal	A Vercel project and token	One call; no app or project

Why does Vercel beat Modal for a waiting agent but not a busy one?

Modal charges the larger of what a sandbox asked for and what it used. A 2 vCPU sandbox is one physical core there, so it pays $0.1419 an hour for that core even while it idles on a model call. Vercel charges only active CPU: $0.128 an hour per CPU, which is dearer than Modal's $0.071 per vCPU, but idle seconds are free.

So the answer turns on how busy the code keeps the CPUs. At this size Vercel stays cheaper until the sandbox averages about 1.2 busy CPUs out of 2, some 72 CPU-seconds in every minute; above that, Modal's flat charge wins.

2 vCPU, 4 GiB, 1,000 one-minute runs	Modal	Vercel	Runtime
20 CPU-seconds of work per run	$3.97	$2.12	$0.64
Both CPUs busy throughout	$3.97	$5.68	$1.33

Runtime measures CPU the way Vercel does, at $0.025 a vCPU-hour, and is the lowest in both rows.

Cost for the same job

Here is the first row worked out: 1,000 runs, 2 vCPU (a single physical core on Modal), 4 GiB of memory (written 4 GB on Vercel), 60 seconds a run and 20 CPU-seconds of actual computing in each.

TextModal    CPU    1,000 × 60 s × 1 core × $0.00003942 = $2.37         Memory 1,000 × 60 s × 4 × $0.00000667      = $1.60         Total                                         $3.97Vercel   CPU    1,000 × 20 s / 3,600 × $0.128       = $0.71         Memory 1,000 × 60 s / 3,600 × 4 × $0.0212  = $1.41         Total                                         $2.12Runtime  CPU    1,000 × 20 s / 3,600 × $0.025       = $0.14         Memory 1,000 × 60 s / 3,600 × 4 × $0.0075  = $0.50         Total                                         $0.64

Over a month of 100,000 runs, Modal comes to $396.60, Vercel to $212.44 and Runtime to $63.89. If both CPUs are busy for the whole minute instead, the job is $1.33 on Runtime, $3.97 on Modal and $5.68 on Vercel. Plan fees, creation charges, disk, network and free allowances are not in these totals.

gVisor or Firecracker?

Modal runs sandboxes under gVisor, a user-space kernel that intercepts the program's system calls. Vercel and Runtime run each sandbox as a Firecracker microVM with its own Linux kernel. Firecracker vs gVisor explains where each draws the boundary.

Which one to pick

• Modal, when the sandboxes need GPUs, or when the team already runs Modal functions, web endpoints and cron jobs and wants sandboxes in the same place. Outside Modal, a sandbox needs a Modal App first.
• Vercel Sandbox, when the product is deployed on Vercel already, or sandboxes must sit in many places: 19 regions, and up to 10,000 at once on Pro.
• Runtime, when the job is an agent that mostly waits: a microVM per sandbox, CPU by use at about a fifth of Vercel's rate, pause and forks that carry memory and live processes, and no app, project or token to create first.

Moving to Runtime

Vercel Sandbox code keeps working with a changed import, and stops needing a Vercel project or token:

TypeScriptimport { Sandbox } from "withruntime/vercel"; // was: from "@vercel/sandbox"

Pythonfrom withruntime.vercel import sandbox  # was: from vercel import sandbox

Modal code is ported call by call. modal.Sandbox.create(app=...) turns into Sandbox.create(), sb.exec(*args) into exec, Sandbox.from_name into Sandbox.getOrCreate(name), and idle_timeout into idlePauseSeconds:

Pythonfrom withruntime import Sandboxwith Sandbox.create(funding="trial") as box:    print(box.exec("python3 -c 'print(6 * 7)'", check=True).stdout)

A coding agent can do either port from the single instruction in migration, working on a branch and checking the result on the free trial. Before the first top-up, npx withruntime switch --from modal or --from vercel matches it with credit, up to $100. The long form is in Runtime vs Modal and Runtime vs Vercel Sandbox.

Sources

Modal and Vercel pricing pages rechecked 25 September 2026; the rest checked 23 September 2026.

• Modal pricing, Modal resources and billing, Modal Sandboxes, Modal security
• Vercel Sandbox pricing, Understanding Vercel Sandboxes
• Runtime pricing

Facts on this page were checked on 25 September 2026.