Agent37 is promoting a pay-as-you-go AI sandbox that costs $4.76 for 730 hours of continuous operation with 2 vCPUs, 4GB of RAM and 4GB of persistent disk. That is an infrastructure estimate, not a subscription price or an all-inclusive budget for running an AI agent.
The more consequential detail is what happens between tasks. Agent37 says compute billing stops when a sandbox sleeps, while files remain available and storage continues billing. For developers assigning a persistent agent to each customer, that could make intermittent usage substantially cheaper than keeping every environment running.
Its pricing comparison lists 31 providers, including Agent37, and advertises costs up to 95% below most alternatives. That is the company’s own comparison of published rates, not an independent performance benchmark. The entry price also comes with an architectural choice: shared host resources and gVisor isolation, rather than a full virtual machine for every sandbox.
What the $4.76 Configuration Actually Includes
Agent37’s published unit rates are $0.80 per vCPU, $0.70 per GB of RAM and $0.09 per GB of disk for a 730-hour reference month, metered by the minute.
The advertised entry configuration therefore breaks down as follows:
- CPU: 2 × $0.80 = $1.60.
- Memory: 4 × $0.70 = $2.80.
- Persistent disk: 4 × $0.09 = $0.36.
- Total for 730 running hours: $4.76.
The CPU and memory portion is $4.40. Storage accounts for the remaining $0.36, assuming the 4GB disk is retained for the entire reference month.
This is the Standard tier, which Agent37 describes as sharing host resources. The pricing page also lists a Performance option at four times the compute rate, with disk pricing unchanged. Developers should therefore avoid treating the lowest advertised price as a promise of dedicated resources or equivalent throughput to another provider’s identically labeled configuration.
Agent37 says billing uses a prepaid balance, with no subscription or plan fee. The advertised $5 starter credit requires a card on file, and the company says nothing is charged until the customer tops up.
Model tokens, search and integrations are separate charges. Developers can bring their own model keys, in which case that model spending does not pass through the Agent37 balance. For an agent making frequent model calls, the $4.76 figure describes only its execution environment.
Sleeping All Month Costs $0.36, Not Zero
The most useful distinction in Agent37’s calculator is between running time and disk retention time.
For the entry configuration, the infrastructure estimate can be expressed as:
$4.40 × running hours ÷ 730 + $0.36 × disk-retention hours ÷ 730
The following calculations use Agent37’s published rates, before credits, taxes and separate model or tool charges:
| Usage pattern | Compute | Retained disk | Total |
|---|---|---|---|
| Runs for 730 hours; disk retained throughout | $4.40 | $0.36 | $4.76 |
| Runs for 57 hours; disk retained for 730 hours | About $0.34 | $0.36 | About $0.70 |
| Sleeps for all 730 hours; disk retained throughout | $0.00 | $0.36 | $0.36 |
| Runs for 57 hours, then is deleted | About $0.34 | About $0.03 | About $0.37 |
That last row explains an important qualification behind the prominently displayed $0.37 estimate. Agent37’s calculator says it assumes the sandbox is deleted after the selected hours. Its 57-hour example therefore includes only 57 hours of storage, rather than retaining the disk across the whole month.
A persistent customer agent that runs for those same 57 hours and sleeps for the rest of a 730-hour month would cost approximately $0.70 instead. Both calculations follow the published rates, but they describe different lifecycle choices.
Agent37’s billing explanation says sleeping or stopping reduces the meter to disk alone, while deletion ends billing. It also advertises automatic sleep when idle and waking on the next request.
Developers should check how that idle policy behaves with their harness. An agent waiting for a model response is not necessarily asleep: Agent37 charges compute while the sandbox runs, rather than only for CPU cycles consumed. Long waits, background processes or a configuration that prevents sleep can move the bill toward the always-on figure.
The 95% Claim Compares Rates, Not Completed Work
Agent37’s comparison is useful as a starting point because it exposes its assumptions and links to the rates behind individual rows. The page marks those provider sources as checked on October 1, 2026.
It is not a standardized test of 30 competing products. The table mixes managed sandboxes, full VMs, container services and infrastructure that developers must configure and operate themselves. Agent37 explicitly labels some AWS and other cloud entries as do-it-yourself options.
The billing models also differ. Agent37’s notes say some providers charge for actual CPU consumption, while its comparison assumes full CPU use where active-CPU billing applies. An agent that spends much of its time waiting on network requests could therefore cost less on those services than the displayed comparison suggests.
Other differences include bundled storage, minimum spending, subscription fees, free allowances and session limits. The page warns that its figures are recurring list-rate estimates, not total invoices or identical hardware, and excludes most competitor storage charges.
The practical question is not simply which row has the lowest hourly equivalent. It is which service completes a representative workload at an acceptable cost, latency and reliability.
A shared-resource sandbox might be inexpensive for a messaging agent that wakes briefly, makes model calls and writes a few files. The same configuration needs closer evaluation for dependency installation, repeated builds or browser-heavy work. Agent37’s price table establishes a low nominal rate; it does not establish how quickly those tasks finish.
gVisor Helps Explain the Price and the Compatibility Limits
In the Hacker News launch post, founder Vishnu described several cost choices: renting hosts from providers such as Hetzner and OVH, optimizing for RAM-heavy agent workloads, and using gVisor to share host resources more efficiently.
According to the founder, agent harnesses often need CPU in bursts and more RAM during browser or computer-use tasks, with demand falling afterward. He says gVisor lets Agent37 use available host memory across sandboxes without reserving a full VM’s memory for each one.
Those are the company’s explanations for its economics, not independently measured utilization results.
The tradeoff is explicit: Agent37’s gVisor sandboxes do not support workloads such as Docker-in-Docker or Android emulators, according to the launch post. The founder says full-VM sandboxes are available for workloads that need them, including running Docker, at a higher price. The cited launch material does not establish an exact full-VM rate, so developers should not apply the $4.76 estimate to that option.
There is also a distinction between deploying a custom Docker image and running Docker inside the resulting sandbox. Agent37 advertises custom images, but that does not remove the founder’s stated nested-Docker limitation.
On security, Agent37 describes its controls as gVisor isolation, separate filesystems, disk quotas and per-sandbox outbound firewalling. Those are vendor statements, not grounds for treating isolation as an absolute guarantee. Developers handling sensitive credentials should evaluate the controls against their own threat model and request supporting security documentation.
Sources
- pricing comparisonagent37.com
- billing explanationagent37.com
- Hacker News launch postnews.ycombinator.com
