Pay by the second.
You set the cap.
Pay-as-you-go, with vCPU and memory metered separately. No idle charge, no platform fee, and a spend cap you set so a runaway swarm can never surprise you.
Hosted Mitos is in early access. Rates here are illustrative; final rates land at general availability.
Estimate assumes each fork holds its full RAM. Copy-on-write sharing usually makes memory less, and forks resume in ~27 ms with no cold-boot tax.
One model: pay-as-you-go, per second
Decoupled meters, so a CPU-heavy and a RAM-heavy sandbox bill differently. You pay only while a sandbox runs, and data transfer out is free.
Illustrative rates, final at general availability. Quoted per hour for readability; metered per second.
How it compares
The two meters every sandbox vendor has, at published rates. Ours are illustrative, but the shape is the point: cheap RAM, and no platform fee underneath.
| Per hour | Mitos* | E2B | Daytona | Modal |
|---|---|---|---|---|
| vCPU / hour | $0.046 | $0.050 | $0.050 | $0.071 |
| RAM / GiB-hour | $0.012 | $0.016 | $0.016 | $0.024 |
Rates are only half of it. Competitors don’t live-fork, so 100 parallel agents run as 100 separate sandboxes, each paying for its own full copy of the base. A Mitos swarm forks one warm parent and shares its memory, so fan-out can cost less than the rates alone suggest, depending on how far each fork diverges.
*Illustrative. Competitor rates published Jun 2026, compute only. E2B ($150/mo) and Modal ($250/mo) also charge a platform fee; Mitos has none.
$100 to start
Free signup credit, no card. Prepaid top-ups after that.
Hard spend caps
Set a ceiling. A runaway swarm suspends at the cap, not at a surprise invoice.
No platform fee
No seat or subscription floor. You pay for compute, nothing else.
No idle charge
Billing is per-second while running. Paused sandboxes cost nothing to run.
Questions worth asking
How is a swarm billed?
Per second, with vCPU and memory metered separately. Forks share the parent’s memory copy-on-write, so fanning out mostly adds vCPU-seconds rather than re-paying for the shared base image. Compute scales with how many subagents run.
Will I get a surprise bill?
No. Hard spend caps are on by default: a runaway loop suspends the org at your limit instead of becoming a five-figure invoice. A soft cap alerts you before that.
Does forking actually make a swarm cheaper?
A fork is a live copy-on-write clone of a running microVM, so you skip the cold-boot tax (it resumes in ~27 ms) and don’t re-build the base image for every subagent. The parent’s pages are shared; each fork adds the pages it writes (about 3 MiB right after a fork, more as it diverges). vCPU still scales with how many subagents run, so a swarm’s bill is mostly compute-time, not duplicated memory. The estimator bills each fork its full RAM, so your real memory cost is usually lower.
Do I pay for idle sandboxes?
No. Compute is per-second while running and scales to zero between forks. You are billed for active work, not for waiting.
Can I run it myself instead?
Yes. The engine is open source under Apache 2.0 and free to run on your own Kubernetes cluster. Hosted Mitos is for teams who want the fork speed without operating the cluster.