In 2026, AI agents clone repos, install dependencies and run tests on their own, many times a day. Every one of those runs needs an isolated machine, so the sandbox has become a real line on the infrastructure bill.
Vercel Sandbox bills you for active CPU time, provisioned memory and each sandbox you create, plus outbound data and storage. Solari bills a sandbox by its size for the time it runs, from $0.0245 per vCPU-hour on Professional plus a per GB-hour memory rate. This guide breaks down each line on the Vercel bill, what the same work costs on Solari, and how to choose between the two.
Why Does Sandbox Pricing Matter Now?
An agent starts far more machines than a person ever did. A developer used to open one environment and keep it all day; a coding agent starts a fresh microVM for each task, runs it for minutes and throws it away.
That changes which part of the bill grows. An agent spends much of each run waiting for a model to answer, so the CPU sits idle while the memory stays allocated. Charge for idle CPU, idle memory or neither, and the same job produces three very different invoices.
Before you compare headline rates, check which billing model matches the way your agent spends its time.
What Does Vercel Sandbox Charge For?
Vercel Sandbox charges for CPU your code actually uses, memory for the whole run, each sandbox you create, outbound data and storage. Each line works differently:
Active CPU. Vercel bills CPU only for the time your code actively uses it. Time spent waiting on network calls, databases or AI model calls does not count, so an agent that waits on its model pays for little CPU. Rates change by region, so price the region your sandboxes run in.
Provisioned memory. Memory is billed per GB-hour for the whole run, idle time included, in one-minute minimum increments. Each vCPU comes with a fixed amount of memory, so a bigger machine always carries more of it. For agents that wait a lot, this is the line that grows.
Creations and network. Every
Sandbox.create()call carries a small fee, priced per million calls, which only shows up at very high volume. Data the sandbox sends out and all traffic on exposed ports is billable. Downloads such as packages and repositories are free.Snapshots and drives. Snapshots and drives bill per GB-month, and snapshots expire after a period without use unless you change the setting. A saved environment that nobody touches can disappear. Set the expiry to match how long your agent needs to come back to its work.
Plans and limits. The free Hobby plan includes a monthly allowance of CPU, memory and creations, and pauses sandbox creation once it runs out. Pro charges usage against a monthly credit and bills the rest at the published rates. Session length, concurrency and maximum machine size all rise with the plan.
Vercel's own worked examples show where the money goes: for every task that mostly waits on a model, from a quick test to a long-running job, memory is the larger part of the estimated cost.
What Does the Same Sandbox Cost on Solari?
Solari bills a sandbox by size, per vCPU-hour and per GB-hour, billed per second while it runs. On Professional, a 4 vCPU, 16 GB sandbox costs $0.221 per hour; on Team, a 4 vCPU, 8 GB sandbox costs $0.114 per hour.
Solari's Professional plan costs $200 a month and the fee comes back as $200 of credits, spent at the rates above. Starter costs $20 a month with $20 of credits. Team lowers the CPU rate to $0.0175 per vCPU-hour, and Enterprise is a custom contract for large fleets. Memory gets cheaper with each plan:
Credits refill each month and there is no end-of-month overage charge: at a zero balance, new sandboxes stop launching until you top up. Snapshots cost $0.05 per GB-month, and the first 10 GB across your organization are free.
How Do the Two Billing Models Compare?
Vercel charges only for active CPU; Solari charges a per-vCPU rate for every second the sandbox runs, and Solari's memory rate is lower than Vercel's on every paid plan. Which bill is smaller depends on how busy your CPU is.
At full-use list price, a 4 vCPU, 8 GB Vercel Sandbox costs $0.683 per hour, against the Solari Team rate above. Vercel bills CPU on active seconds only, so that figure is the price of a sandbox that keeps every core busy, such as a build or a test suite.
An agent that idles on model calls pays Vercel far less CPU than that, and memory becomes most of its bill. Both platforms bill memory for the whole run, so compare the GB-hour rates first, then the CPU rates at the utilization your agent actually reaches.
When your agent goes quiet, a Solari sandbox pauses after its idle timeout by default and resumes exactly where it stopped. A long gap between tasks never means rebuilding the environment.
What Role Does Solari Play?
Solari runs sandboxes, cloud browsers and desktops for AI agents from one API, so the code your agent writes and the websites it operates bill to one account.
Solari cloud browsers cost $0.07 per browser hour on Professional, from the same credit balance as your sandboxes. Your agent drives them with Playwright or any CDP client. A coding agent that also checks a deployed page, signs in to a dashboard or fills a form does it without a second vendor.
Solari supports up to 100,000 concurrent sandboxes on Enterprise, so you can run a whole evaluation set or a large batch of tasks at once. Small teams start with 10 sandboxes on Professional and grow from there.
Team and Enterprise have unlimited session duration, so a long-running agent is never cut off by a clock. Professional allows 24-hour sessions.
Each dedicated Solari sandbox is a microVM with a dedicated kernel per session, run by Cloud Hypervisor, an open-source virtual machine monitor. Your agent's untrusted code never shares a VM with another organization.
You can take a snapshot of a running sandbox, revert it after a bad step or fork it to try two approaches in parallel. Your agent keeps a known good state instead of rebuilding it every run.
For computer use models, Solari desktops give an agent a full Linux screen at the same per-size rates, plus $0.02 an hour for the live screen. The introduction to Solari covers how the three surfaces fit together.
Where Is Sandbox Pricing Going Next?
Sandbox billing is moving closer to the way agents actually work. Over the next year, expect four changes:
Billing that follows real work. More vendors will separate busy time from idle time, the way Vercel does with active CPU. Memory, not CPU, becomes the line teams tune.
Pause as a cost habit. Agents will pause and resume sandboxes between steps instead of holding them open. Lifecycle settings will matter as much to the bill as the hourly rate.
Contracts for fleets. As agent fleets grow, teams will buy sandbox capacity the way they buy cloud compute, on volume contracts. List prices will matter mainly for the first few months.
One account for every surface. Agents that write code, use browsers and operate desktops will push teams toward one vendor and one bill. Splitting the stack across three invoices will look like extra work.
What stays human is the judgment about what runs: which tasks deserve a large machine, how long an environment should live, and which agent output is safe to ship.
Get Started With Solari
Start with the build-and-test loop your coding agent runs most, run it on a 4 vCPU sandbox for a week and read the real bill. Vercel is the more established platform with a large developer ecosystem; Solari is the younger company and moves faster on speed and price. Solari is the fastest and cheapest way to run cloud browsers and VMs for coding agents. Get started with a free account and start running your agents today.
