Cheapest Sandbox for AI Agents: 15 Providers Compared (2026)

Do not index

Agent37 has the lowest recurring list-rate estimate in this comparison for a standard sandbox with 2 vCPU, 4 GB RAM and 730 running hours: $4.76, including 4 GB of persistent disk. If you are looking for the cheapest sandbox for AI agents that keep working throughout the month, that is a useful starting point.

There are cheaper ways to run a few short jobs, especially inside another provider's free allowance. This guide shows where the price advantage holds, what each estimate leaves out, and how to calculate your own workload.

We build Agent37. We checked the providers' public pricing pages on October 1, 2026 and linked the sources below. These are price calculations, not performance benchmarks or quotes for identical hardware.

Which is the cheapest sandbox for AI agents?

For the standard 2 vCPU / 4 GB scenario below, Agent37 costs $4.76 for 730 running hours. Northflank's published compute rates produce $48.66, Morph's standard compute-unit rate produces $73.00, and E2B and Daytona each produce $120.89 before any plan fees or additional storage.

Agent37 bills by the minute from a prepaid balance. The monthly number is an estimate for continuous uptime, not a subscription. A sandbox that sleeps between tasks pays less compute while retaining its files.

Calculate your sandbox cost and compare providers.

The comparison: 15 providers at the same running time

The baseline is 2 vCPU, 4 GB RAM and 730 aggregate running hours per month. Agent37's estimate includes 4 GB disk retained all month. Most competing rows show compute charges; storage is discussed separately. Hetzner includes its bundled disk. Cloudflare includes runtime disk and needs 6 GiB RAM to meet its custom-size rules. Blaxel matches 4 GB RAM with automatically allocated CPU.

For active-CPU billing, the baseline assumes both CPUs stay busy throughout runtime. Sprites also assumes 4 GB actual memory usage. Modal uses the requested resource amount. Less CPU activity can reduce the bill on providers that charge for actual consumption.

Free credits, recurring allowances, subscriptions, taxes and network charges are excluded. A monthly estimate does not mean every provider permits a single uninterrupted month-long session.

730 running hours, 15 providers

Nominal 2 vCPU / 4 GB RAM. Full CPU use assumed. USD before credits, plan fees and most competitor storage. Cloudflare uses 6 GiB; Blaxel matches RAM.

Agent37$4.76

Hetzner$6.49

Northflank$48.66

AWS EC2$55.85

Morph$73.00

Fly.io Sprites$119.72

Daytona$120.89

E2B$120.89

Blaxel$120.89

Hopx$120.89

Cloudflare Sandbox$145.28

Beam$165.83

Modal$173.71

Runloop$231.26

Vercel Sandbox$248.78

Agent37 includes 4 GB retained disk. Hetzner is retained all month; AWS and Hetzner are DIY VMs. Checked October 1, 2026. Prices compare billing, not hardware performance.

Provider	730-hour usage estimate	62-hour usage estimate	What the number represents
Agent37	$4.76	$0.73	Standard compute plus 4 GB retained disk
Hetzner	$6.49	$6.49	CX23 VM retained all month, EU; DIY setup
Northflank	$48.66	$4.13	Compute at infrastructure unit rates
AWS EC2	$55.85	$4.74	Linux c6a.large in us-east-1; DIY setup
Morph	$73.00	$6.20	Two MCUs/hour at the standard $0.05 rate
Fly.io Sprites	$119.72	$10.17	Two active CPUs and 4 GB actual memory
Daytona	$120.89	$10.27	Allocated CPU and RAM
E2B	$120.89	$10.27	CPU and RAM, before plan fees
Blaxel	$120.89	$10.27	4 GB RAM with automatic CPU allocation
Hopx	$120.89	$10.27	CPU and RAM, before disk
Cloudflare Sandbox	$145.28	$12.34	Containers: 2 vCPU, 6 GiB RAM, 4 GB runtime disk
Beam	$165.83	$14.08	Sandbox compute, using one physical core
Modal	$173.71	$14.75	One requested physical core and 4 GiB RAM
Runloop	$231.26	$19.64	Devbox compute, before storage and plan fees
Vercel Sandbox	$248.78	$21.13	Full active CPU at the displayed base regional rate

The 62-hour column models roughly two hours of running time per day while keeping an Agent37 sandbox's disk for the month. It does not charge sleeping compute as if it were still running. GB and GiB follow each provider's published billing units, so the memory quantities are nominal comparisons.

How Agent37's $4.76 price works

The standard monthly rates are $0.80 per vCPU, $0.70 per GB RAM and $0.09 per GB disk, using a 730-hour month. The entry shape works out to:

CPU: 2 × $0.80 = $1.60.

RAM: 4 × $0.70 = $2.80.

Persistent disk: 4 × $0.09 = $0.36.

Total: $4.76 for 730 running hours.

For a sandbox that runs for 62 hours, compute is $4.40 × 62 ÷ 730. Add $0.36 for retained disk and the total rounds to $0.73. If it stays stopped all month, the same disk is $0.36. Delete the sandbox and its meter stops.

Standard instances share host resources. Performance instances receive higher CPU priority under contention and cost four times the compute rate, with unchanged disk pricing. The entry performance shape is therefore $17.96 for 730 hours. The pricing calculator lets you switch between both.

Hosting does not include unlimited AI inference. Model tokens, search and integrations are separate usage charges, or you can bring your own provider keys. The instance documentation explains resource sizes and lifecycle controls.

Provider rates and the costs behind the bars

Northflank: the closest managed compute price in this list

Northflank pricing lists $0.01667 per vCPU-hour and $0.00833 per GB-hour. For two CPUs and 4 GB, multiply $0.06666 by 730 to get $48.66. Persistent SSD storage adds $0.15 per GB-month; egress is separate. This row uses infrastructure unit rates, so check the plan and configuration you actually deploy.

Morph: bundled resources through compute units

Morph pricing uses compute units at a standard rate of $0.05 each. Hourly units are the largest of vCPU count, RAM divided by 4 GiB rounded up, and disk divided by 16 GiB rounded up. The baseline needs two units per hour: $73.00 across 730 hours. Paid plans discount purchased units. Snapshot retention has its own cost.

E2B and Daytona: equal compute rates, different packaging

E2B pricing lists $0.000014 per vCPU-second and $0.0000045 per GiB-second. That produces $120.89. Hobby has one-hour sessions and 10 GiB storage. Pro adds a $150 monthly subscription, custom CPU/RAM and sessions up to 24 hours, with 20 GiB storage. The chart excludes that subscription; it is not an E2B Pro invoice.

Daytona pricing lists the equivalent hourly rates: $0.0504 per vCPU and $0.0162 per GiB RAM. Compute is also $120.89. The first 5 GiB disk is free, with additional storage at $0.000108 per GiB-hour. Both providers offer trial credits, which can cover early experimentation.

Blaxel and Hopx: another two $120.89 compute estimates

Blaxel pricing bills sandbox active runtime at $0.0000115 per GB RAM-second. Four GB produces $120.89, with CPU scaling automatically. This is a memory match, not a promise of exactly two vCPUs. Snapshot storage, persistent volumes and images cost extra. Its suspension behavior can reduce active time for intermittent work.

Hopx pricing lists the same CPU and memory rates as E2B, giving $120.89 before disk. Storage adds $0.00000003 per GiB-second. Use the individual rate lines to reproduce the calculation instead of relying on a rounded example workload.

Fly.io Sprites and Cloudflare: CPU activity matters

Sprites pricing charges $0.03825 per actual CPU-hour and $0.021875 per actual GB-memory-hour. The fully used baseline is $119.72. Storage is separate: $0.000683 per hot GB-hour and $0.000027 per cold GB-hour. A process that mostly waits and uses little RAM can cost much less than this row.

Cloudflare Containers pricing charges active CPU, provisioned memory and runtime disk. Its custom-size limits require at least 3 GiB per vCPU, so two vCPUs need 6 GiB. At full CPU use with 4 GB runtime disk, the estimate is $145.28. The $5 Workers plan, included allowances, Workers and Durable Objects usage, and persistent storage are outside this estimate. This is the Containers-backed sandbox configuration.

Beam and Modal: convert physical cores correctly

Beam's sandbox rates are $0.0000375 per physical-core-second and $0.0000064 per GiB-second. One physical core represents two vCPUs, producing $165.83. Beam also lists cheaper serverless-function rates; those are a different product and are not used here.

Modal sandbox pricing uses $0.00003942 per physical-core-second and $0.00000667 per GiB-second, producing $173.71. Modal bills the greater of requested and actual resources. The calculation requests the full baseline. Monthly compute allowances can reduce out-of-pocket costs, particularly for small workloads.

Runloop and Vercel: check subscriptions and CPU billing

Runloop pricing lists $0.108 per CPU-hour and $0.0252 per GB-hour: $231.26 for baseline compute. Basic has no subscription; Pro adds $250/month and includes suspend/resume. Storage charges and allowances are separate from the chart.

Vercel Sandbox pricing shows a base regional rate of $0.128 per active CPU-hour and $0.0212 per GB-hour of provisioned memory. That gives $248.78 at full CPU use. Waiting on I/O reduces CPU charges. Included usage, subscription fees, creations, storage and transfer also affect the bill; region changes can change rates.

Hetzner and AWS EC2: inexpensive machines you operate

Hetzner's published EU prices put CX23 at $6.49/month, excluding IPv4 and VAT. The chart retains the VM all month, even when powered off. Deleting it earlier can reduce the bill. It includes more disk than the Agent37 entry shape, so storage-heavy workloads deserve a separate comparison.

AWS EC2 on-demand pricing puts Linux c6a.large in US East at $0.0765/hour, or $55.85 for 730 hours. EBS storage, public IPv4 and transfer are additional. These two rows are DIY VM alternatives: installation, sandbox isolation, patching and backups are work you must arrange.

When the cheapest option changes

A free allowance can beat a paid usage estimate. E2B and Daytona offer trial credits; Modal and Vercel include usage allowances. If your entire workload fits an allowance, your immediate bill may be zero. Compare the cost after credits expire and the restrictions attached to the plan.

A one-hour job is different from a persistent agent. Keeping Agent37's 4 GB disk for a month and running for one hour costs about $0.37. Northflank's one-hour baseline compute is about $0.07 before storage. A job that needs no retained state can be cheaper elsewhere. Deleting an Agent37 sandbox also ends disk billing rather than forcing a full-month charge.

CPU capacity is not a performance benchmark. Shared resources, physical cores, automatically allocated CPU and actual-use billing are different. Test representative agent turns, repository installs and browser tasks before deciding by the bar lengths alone.

Persistence has value only when you need it. A customer agent with files, settings and recurring jobs benefits from keeping its environment. A disposable evaluation may prioritize startup latency, concurrency or snapshot branching. Those requirements can matter more than the lowest ongoing hosting rate.

Calculate your real agent budget

Start with three numbers: how many environments you retain, how many hours each runs, and how much disk each needs. Then add model tokens, tool calls and any provider-specific fees.

For example, 100 standard entry Agent37 sandboxes running for 730 hours cost $476 in sandbox hosting. That excludes AI usage. At the same aggregate compute time, E2B's listed CPU and RAM rates produce $12,088.80 before subscriptions. This is capacity arithmetic, not a promise that both setups deliver equal throughput.

Stop or sleep instances between tasks when the workflow permits it. An awake but quiet Agent37 sandbox still pays compute. Use a separate model budget so a cheap sandbox does not hide an expensive inference loop. Our guide to agent running costs explains that distinction in more detail.

Frequently asked questions

Is Agent37 always the cheapest sandbox?

No. It has the lowest standard 730-hour managed estimate in this comparison. Free allowances, short jobs, lower actual resource use and negotiated prices can change the result. The table states the workload so you can decide whether it fits yours.

Does the $4.76 price include an AI model?

It covers the standard sandbox and 4 GB persistent disk. Model tokens and tools are separate, whether you use managed credentials or bring your own keys.

Can I try Agent37 before paying for compute?

New eligible accounts receive $5 starter credit after adding a card. Adding the card does not charge it. The credit is enough for a little over 730 hours of entry sandbox hosting if you spend nothing on models or tools. It is a one-time credit, not an ongoing free Cloud tier.

What should I use for an agent in my product?

If each customer needs a persistent environment and ongoing compute cost matters, try the standard Agent37 shape with a representative workload. The Cloud API lets you provision and control instances programmatically, while the pricing page shows what changes as runtime and size increase.

Start with $5 Cloud credit, or compare your workload on the pricing page.