Agent37 vs Vercel Sandbox vs E2B: 100 Agent-Hours Priced
A developer measured a coding agent at 24.6% of two vCPUs, five times the figure in the Vercel pricing fight, then repriced all three sandboxes.

A coding agent burns a median 24.6% of two vCPUs — five times the figure at the centre of a public sandbox pricing fight.
That number had never been measured in public, and every price in the argument depends on it.
The fight started on 3 November, when someone posted a price list for one always-on agent sandbox with 2 vCPUs and 4 GB of RAM: Vercel $249 a month against Agent37's $4.76. Vercel CEO Guillermo Rauch replied that it was "a total miscalculation of active cpu time" behind "fake cost comparisons" — Vercel Sandbox bills CPU only while code is actually running, so an idle sandbox does not cost $249. On 16 November Agent37 answered with new math: 100 coding agents for an hour at "just 5% active CPU", which its video priced at $0.68 on Agent37 against $9.75 on Vercel.
Nobody had measured the workload
The hands-on writeup measured it, with one limit — it never touched Agent37's cloud. The author pulled the public opencode-base image, ran it under Docker at the same 2 vCPU / 4 GB limits as the default shape, and read CPU and memory from each container's cgroup. Create times, wake-from-sleep and shared-host behaviour went untested.
What came back: a median of 24.6% of the 2-vCPU shape, ranging 19.2% to 29.6%, with one-second peaks up to 1.59 vCPU during npm install and test runs. Peak process memory ran 755–873 MB, well inside 4 GB. The agent was rarely idle even while waiting on the model — only 37 of 384 one-second samples taken during turns were under 0.05 vCPU, because the process streams output, parses tool calls and writes files throughout.
At the measured median of 15.2 CPU-seconds per turn, the 5% figure works out to about 24 turns an hour — one every 2.5 minutes. That describes an agent pairing with a human who reads each answer. It does not describe agents running back to back on their own.
The three-way bill for 100 sandbox-hours
Prices were taken from public rate cards on 17 November: Agent37's billing docs, Vercel Sandbox Pro in iad1 ($0.128 per active vCPU-hour plus $0.0212 per GB-hour of provisioned memory), and E2B ($0.000014 per vCPU-second plus $0.0000045 per GiB-second).
| CPU assumption | Agent37 | Vercel Sandbox | E2B |
|---|---|---|---|
| 5% active (the thread's number) | $0.65 | $9.76 | $16.56 |
| 24.6% active (measured) | $0.65 | $14.77 | $16.56 |
Two things fall out of that table. First, memory, not CPU, is most of Vercel's bill: at 5% active CPU, 87% of the Vercel price is provisioned RAM, and even a fully idle Vercel sandbox costs $0.085 an hour — 13 times Agent37's flat rate. Second, a busier agent makes Vercel more expensive, not cheaper; at 24.6% its price rises about 51% over the 5% row. Agent37's guaranteed-core tiers would cost $1.56 (vm) and $2.46 (performance) for the same 100 sandbox-hours.
What the hourly rate does not cover
Three caveats matter more than the headline number. Agent37's $4.76 shape runs on shared hosts, and its own docs say sustained heavy CPU "may be limited when the hardware is busy" — whether those 1.6 vCPU install bursts get throttled is exactly what a local test cannot answer. Concurrency is gated by spend: 100 agents at 4 GB each means 400 GB awake, and Agent37 allows 32 GB before your first top-up, 200 GB at $50 of top-ups and 800 GB at $250, against 10,000 concurrent sandboxes on Vercel Pro and 10 (Hobby) or 100 (Pro, $150/month) on E2B. And the models differ: Agent37 instances live until deleted and bill disk alone while asleep, while Vercel and E2B cap a session at 24 hours (1 hour on E2B Hobby).
What to check before you move
If your agents are long-lived per-user computers that idle between turns, the persistent-VPS model wins on price by an order of magnitude, and the active-CPU argument cannot close a gap that is mostly RAM. If they are throwaway jobs at high concurrency, the quota ladder — not the hourly rate — is the number to price. Either way, measure your own agent's CPU per turn before trusting any comparison table, including this one.
More from DangMua