How the costs are calculated

Every provider fits one of these, sometimes with a twist.

• Per resource, on what you reserve. CPU, memory and sometimes disk each have an hourly rate, charged for the size you asked for whether or not it is busy. E2B, Daytona, InstaBox, Agent 37, Modal and boxd's CPU work like this.
• Per resource, on what you use (active-CPU billingThe processor is charged only while it is working, not while it waits. Cheap for agents that spend most of their time idle.). CPU is charged only while it works, so an idle agent pays mostly for memory. Runtime (with a small floor), Vercel Sandbox, Cloudflare Sandbox and Sprites work like this, and boxd charges memory on what is actually resident.
• A flat price per size. You pick from a short list of sizes, each with one hourly price. Boat, Sandboxer and Lizard work like this. If no size matches, you buy the next one up.
• Compute units. Morph charges one price per unit-hour, where the number of units is set by whichever of CPU, memory or disk is largest relative to the others.

On top of usage sits the plan. Some plans are a fee that comes back as usage (Boat, Sandboxer, Lizard, InstaBox, Vercel Pro), some add usage on top of a fee (E2B Pro, Cloudflare's Workers Paid minimum), some include a monthly credit (Modal) and some have no plan at all.

1. Running hours. Tasks: tasks × minutes ÷ 60. Machines: machines × hours a day × days, or 730 for always-on. For providers that sleep idle machines, running hours are multiplied by the awake share and the rest counts as stopped time.
2. Size. Where a provider sells fixed sizes, the cheapest size with at least the vCPU and memory you asked for is used, counting any volume needed for extra disk. The size used is printed with every result.
3. Compute. Reserved CPU is running hours × vCPU × rate. Used CPU is running hours × the larger of (vCPU × CPU working) and the provider's floor × rate. Memory is running hours × GiB × rate.
4. Disk. Charged at the provider's running rate while running and its stopped rate while stopped, after any free allowance. Where the provider prices in decimal gigabytes, GiB are converted (1 GiB = 1.074 GB).
5. Plan. Each plan is priced: its fee, plus usage beyond what it includes (rounded up to whole packs where overage is sold in packs). Free monthly allowances are subtracted at the provider's own rates. The plan with the fewest problems wins, then the cheapest.
6. Problems. A result is marked needs a change when the size does not exist, the disk is over a limit, a session limit is shorter than your sessions, or stopping would delete files you asked to keep. The figure is still shown, for the nearest thing on sale.
7. Unknowns stay unknown. A price a provider does not publish is not guessed; it is named in the notes and left out of the total.

• Model tokens, unless you add them. For most agents the language model costs more than the machine. The calculator can add an estimate: messages a day × thousands of tokens each × 30 days, priced at the median of the model hosts that keep no data (zero data retention), from TokenWatch. It assumes TokenWatch's agent mix of 5% new input, 90% cached input and 5% output, and it is the same on every sandbox provider, so it never changes the order.
• Egress, public IP addresses, seats and taxes. Where a provider charges them, its notes say so.
• Snapshots and backups whose size or price is not published.
• One-time sign-up credits. They are mentioned, not subtracted, because they run out. Recurring monthly credits are subtracted.
• Concurrency limits and start limits. A thousand tasks in one burst can hit limits that the same tasks spread over a month never meet.
• Discounts, commitments and enterprise contracts.

battleships.dev, built by the Boat team, is an estimator that prices one workload across several hundred providers. The calculator here borrows its idea of describing a workload by size, sessions and how busy the machine is, and its rule that an unpublished price is unknown rather than free. Thanks to them for publishing their data openly.

The differences are deliberate:

• Every provider checked by hand. Every rate here comes from the provider's own pricing page or documentation, with a date, rather than from Battleships' cards. Battleships' code is not used.
• Where we disagree, we say so. Modal documents one physical core as two vCPUs, so its CPU here costs half of what Battleships charges per vCPU. Battleships lists Boat's disk sizes as 12, 50, 125 and 251 GB; Boat's own pricing page says 51, 51, 111 and 251 GB, and those are used here. Other differences are listed on each provider's page.
• Fewer inputs. Battleships also models concurrency, seats, egress, snapshots and network access tiers. This calculator keeps to the inputs a first-time reader can answer, and names what it leaves out.
• The answer is the start, not the end. Each result links to the provider's page and to the features that decide whether it fits, because the cheapest machine that cannot keep your files is not cheap.