Summary: An AI agent sandbox is billed for very different things depending on the provider: Agent37 charges a flat per-minute rate for the whole time a sandbox is awake, Vercel Sandbox charges memory for the whole run but CPU only while your code actually computes, and E2B charges CPU and RAM for every running second. I priced the same 2 vCPU / 4 GB sandbox on all three from their own pricing pages and ran three realistic agent workloads through the numbers. The answer depends far more on how much your sandbox sits idle than on the headline rate.
Why sandbox pricing is confusing right now
Coding and computer-use agents need somewhere safe to run commands, install packages and execute the code they write. That has turned "agent sandboxes" into a crowded category, and the providers price them in ways that are hard to line up side by side. One quotes a monthly price per vCPU, another a per-second rate, another a per-hour rate that only counts CPU time you actually burn.
Agent37's pricing page publishes a 31-provider comparison that puts itself at the bottom. That table is useful, but it is the vendor's own framing: its fine print says it assumes full CPU use wherever a provider meters active CPU. Agent workloads rarely look like that. Most of an agent's wall-clock time is spent waiting for a model to answer, and on a provider that only bills active CPU, that waiting is nearly free. So I rebuilt the math myself from each provider's official rates and tested it against idle-heavy workloads.
The published rates (checked 2026-11-18)
Agent37
- Standard compute: $0.80 per vCPU and $0.70 per GB of RAM per 730-hour month, which the pricing page also shows as $0.0011 per vCPU-hour and $0.00096 per GB-hour.
- Disk: $0.09 per GB per month ($0.00012 per GB-hour), billed while you keep the sandbox, even when it sleeps.
- Metering: per minute. Running and transitional states (starting, restarting, updating, stopping) bill the full compute rate; sleeping, stopped or waking bill disk only; cold storage is $0.03 per GB-month on compressed bytes.
- Isolation tiers: Standard shares host resources. A "VM" instance (own kernel) costs 2.5x compute, and a "Performance" instance (dedicated cores) costs 4x compute. Disk is unchanged.
- Credit: $5 starter credit after you add a card; no subscription.
Vercel Sandbox
- Active CPU: $0.128 per hour (iad1 region), counted only while code uses the CPU. Waiting on network, databases or model calls does not count.
- Provisioned memory: $0.0212 per GB-hour for the whole time the sandbox runs, idle included, in 1-minute minimum increments. Each vCPU comes with 2 GB.
- Creations: $0.60 per million.
- Hobby (included): 5 Active CPU hours, 420 GB-hours of memory and 5,000 creations per month, 45-minute sessions, 10 concurrent sandboxes. Pro usage is charged against the plan's $20/month credit first, with 24-hour sessions.
- Rates for CPU, memory, transfer and drives vary by region.
E2B
- CPU: $0.000014 per vCPU-second ($0.0504 per vCPU-hour).
- RAM: $0.0000045 per GiB-second ($0.0162 per GiB-hour).
- Hobby: free, with a $100 one-time usage credit, 1-hour sessions and 10 concurrent sandboxes; no credit card required.
- Pro: $150/month plus usage, 24-hour sessions, 100 concurrent sandboxes (expandable to 1,100), custom CPU and RAM. Pro adds no extra credits.
- Paused sandboxes stop compute billing; E2B's docs say you pay only while a sandbox is running.
Three agent workloads, priced
All three use the same shape: 2 vCPU, 4 GB RAM (4 GiB on E2B) and, for Agent37, 4 GB of disk. For Vercel I show two columns, because its bill swings with CPU utilization. Vercel's own example calculations assume 10% utilization "reflecting time spent waiting for LLM calls", so that is the realistic column; the 100% column is what you pay if the sandbox is busy compiling or running tests the whole time.
| Workload (per month) | Agent37 Standard | Agent37 VM | Vercel Sandbox, 10% CPU | Vercel Sandbox, 100% CPU | E2B |
|---|---|---|---|---|---|
| A. Always on, 730 h | $4.76 | $11.36 | $80.59 | $248.78 | $120.89 |
| B. Workday agent, 8 h x 22 days (176 h awake, asleep otherwise) | $1.42 | $3.01 | $19.43 | $59.98 | $29.15 |
| C. 300 short tasks x 5 min (25 h), deleted after each | $0.16 | $0.39 | $2.76 (inside Hobby quota) | $8.52 | $4.14 |
| Free allowance | $5 starter credit (card required) | Hobby: 5 CPU-h, 420 GB-h, 5,000 creations / month | $100 one-time credit | ||
| Max session | Auto-sleep and wake; no session cap stated on the billing page | 45 min Hobby / 24 h Pro | 1 h Hobby / 24 h Pro | ||
Usage charges only, at list rates. Plan fees (E2B Pro $150, Vercel Pro), taxes, network transfer, snapshots and storage beyond what is stated are excluded. Agent37 rows A and B include $0.36 of disk kept all month.
How I got these numbers
- Agent37 Standard, workload B: 176 h x (2 x $0.0011 + 4 x $0.00096) = $1.06 compute, plus 4 GB x $0.09 = $0.36 disk for the month.
- Vercel, workload B at 10%: 176 h x (2 vCPU x 0.10 x $0.128 + 4 GB x $0.0212) = $19.43. Memory is about 77% of that bill.
- E2B, workload B: 176 h x 3,600 s x (2 x $0.000014 + 4 x $0.0000045) = $29.15.
- Vercel, workload C: 25 h x 2 vCPU x 10% = exactly 5 Active CPU hours and 100 GB-hours of memory, so it lands right at the Hobby limit. A little more CPU and Hobby pauses sandbox creation for the rest of the cycle.
You can check the arithmetic against the providers' own calculators: Agent37's estimate for 57 hours of 2 vCPU / 4 GB is $0.37, and Vercel's documented "AI code validation" example (5 minutes, 2 vCPU, 4 GB, 10% CPU) comes to $0.0092, both of which match the formulas above.
What the numbers actually say
1. Agent37 is cheapest on paper by a wide margin
Even with the 2.5x VM multiplier, Agent37 stays far below the other two in every workload. If list price per awake hour is your main constraint and Standard's shared-host isolation is acceptable for your code, it is hard to beat. Its "94% lower cost" headline comes from the vendor's own comparison, though, so treat that figure as a claim and compare the like-for-like tier you would actually run.
2. Vercel's gap shrinks a lot for waiting-heavy agents
Agent37's comparison prices Vercel at full CPU. At a realistic 10% utilization, Vercel's bill is about a third of that, and for small bursty jobs the Hobby quota can cover the whole month. The catch is memory: it is billed for every running minute, so a large RAM allocation left open all day is what drives a Vercel bill up, not CPU.
3. E2B's cost is all about pausing
E2B bills CPU and RAM per running second regardless of utilization, so an always-on sandbox is the most expensive pattern there. Its docs push auto-pause, sbx.pause() and sbx.kill() for exactly this reason. For short tasks, the $100 one-time credit covers a lot of runs before you pay anything, and Hobby needs no card.
4. Plan fees can matter more than usage
If you need sessions longer than an hour or custom CPU/RAM on E2B, Pro is $150/month before usage, which dwarfs the usage figures for workloads B and C. On Vercel, Pro usage first draws down the plan's $20 monthly credit. On Agent37 there is no plan fee, but the wallet has to stay funded.
Editor's note
- Agent37 fits long-lived personal or team agents that sleep a lot and where low cost per awake hour matters most. Not ideal if you need strong isolation but only budget for Standard; price the VM or Performance tier instead.
- Vercel Sandbox fits apps already on Vercel and agents that mostly wait on model calls. Watch the default 5-minute timeout (extend it with
sandbox.extendTimeout()) and the 45-minute Hobby session cap. - E2B fits high-volume short code-execution jobs and teams that want per-second billing with pause/resume. Watch the 1-hour Hobby session cap; Pro raises it to 24 hours.
- Common error on Agent37: creating an instance returns
402 insufficient_balancewhen the wallet holds less than one day of that instance's running rate. Top up first; nothing is debited at create, so a failed create costs nothing. - Common surprise on Vercel Hobby: once a quota is exceeded, sandbox creation pauses until the cycle resets. You are not charged, but your agent stops getting sandboxes.
- Common surprise on E2B: when free credits run out and no payment method is on file, projects in the workspace are blocked shortly afterwards. Set a spending limit in the console if you add a card.
- Another Agent37 error:
409 instance_limit_reachedmeans the awake instances in your workspace would exceed its memory limit (32 GB until your first top-up). Sleeping instances do not count toward it. - Agent37 keeps billing disk while a sandbox sleeps (its docs put a mostly-asleep 4 GB instance at about $0.36 a month); delete sandboxes you no longer need rather than letting them sit.
Which one to pick
- Estimate awake hours, not task count. Multiply the hours your sandbox stays running by the per-hour rates above; that number dominates every bill.
- Estimate CPU utilization. If your agent mostly waits on an LLM, Vercel's Active CPU model rewards you. If it compiles and tests constantly, use the 100% column.
- Match isolation tiers. Compare Agent37 VM, not Standard, if you need a full VM like the others provide.
- Add the plan fee for the session length and concurrency you need.
- Start inside a free allowance (Vercel Hobby, E2B's $100 credit or Agent37's $5 credit) and read the provider's usage page after a real week before committing.
FAQ
Is Agent37 really 94% cheaper than other sandboxes?
That is Agent37's own claim, based on its comparison assumptions (including full CPU use for providers that meter active CPU). In my three workloads its Standard tier was still the cheapest, but the gap to Vercel shrinks considerably at realistic agent CPU utilization.
Does a paused or sleeping sandbox cost anything?
On Agent37, a sleeping sandbox bills disk only ($0.09 per GB-month). On E2B, compute billing stops when a sandbox is paused, killed or times out. On Vercel, provisioned memory and Active CPU are billed only while the sandbox runs; snapshots and drives have their own storage rates.
Can I run an agent sandbox for free?
For light use, yes: Vercel Hobby includes 5 Active CPU hours and 420 GB-hours of memory a month, and E2B Hobby includes a $100 one-time credit. Agent37 gives a $5 starter credit once you add a card. Session caps (45 minutes and 1 hour) apply on the free tiers.
Related on lightmind24
- Modal vs Runpod vs Lambda: What an H100 Hour Really Costs
- AWS Emulator Tutorial: Let Coding Agents Test AWS Locally
- Codex CLI Not Working? Fixes for Login, Sandbox and Limits
Cheap sandboxes are easy to find; the real saving comes from letting yours sleep when your agent is just waiting.
Official screenshots of the Agent37, Vercel Sandbox and E2B pricing pages, plus my own cost table, are at the end of this post. Image credits: screenshots of agent37.com, vercel.com and e2b.dev pricing and docs pages; cost table by lightmind24.
lightmind24 is an independent review site, not affiliated with Agent37, Vercel or E2B.
Last verified: 2026-11-18
Sources:
Agent37 pricing
Agent37 billing docs
Vercel Sandbox pricing and quotas
E2B pricing
E2B billing and limits docs
댓글
댓글 쓰기