AI AGENTS · PRICING COMPARE
Agent sandbox pricing compare prices verified 2026-08-25
Capture-dated sandbox rate cards, ranked on your actual workload. Vendors without verifiable public rates stay unranked instead of guessed onto the chart.
Workload presets
Every figure above comes from the same ranked rows below — same inputs, same math, no hand-typed numbers.
Interactive Cost Breakdown
Compute vs memory share of each computed bill at your current inputs — for the vendors matching your search/filter above. Bar lengths scale to the most expensive visible vendor.
Northflank
- vCPU
- $0.02/hr
- RAM
- $0.0083/GiB-hr
- Billing
- per-second
Daytona
- vCPU
- $0.05/hr
- RAM
- $0.02/GiB-hr
- Billing
- per-second
- Credit
- $200.00 free
E2B
- vCPU
- $0.05/hr
- RAM
- $0.02/GiB-hr
- Billing
- per-second
- Credit
- $100.00 free
Cloudflare Containers
- vCPU
- $0.07/hr
- RAM
- $0.009/GiB-hr
- Billing
- per 0.01s
Beam (sandbox)
- vCPU
- $0.07/hr
- RAM
- $0.02/GiB-hr
- Billing
- per 0.001s
- Credit
- $30.00 free
Modal (Sandboxes)
- vCPU
- $0.07/hr
- RAM
- $0.02/GiB-hr
- Billing
- per-second
- Credit
- $30.00 free
Self-host Firecracker (Hetzner)
- vCPU
- —
- RAM
- incl.
- Billing
- per nulls
- Self-host
- $153.59/mo host
Runloop
- vCPU
- $0.11/hr
- RAM
- $0.03/GiB-hr
- Billing
- per-second
- Credit
- $50.00 free
Why not just a VPS?
A VPS is one long-lived box you administer. An agent sandbox is disposable compute spawned per task by an agent loop — potentially hundreds of times a day, living seconds to hours, then destroyed. Four capabilities you're buying over a raw box:
- Snapshot cold starts. Pause the entire filesystem + process state mid-task and resume later where the agent left off, in about a second. A VPS boots from disk image; it can't freeze a running session and unfreeze it on another machine.
- Per-second billing, zero commitment. No monthly plan — idle costs nothing because nothing exists when no task is running. A VPS bills whether your agent used it today or not.
- Isolation for untrusted code. You're running AI-generated code nobody has reviewed. A microVM blast radius matters at 500 spawns a day in a way one SSH box never confronts.
- Spawn/exec SDKs instead of SSH. Built for a program calling create → run → destroy programmatically, with structured output — not for a human sysadmin.
You can roll your own: the DIY row above prices Firecracker microVMs on a plain VPS (Hetzner CPX62), which wins at high sustained volume — but then you own image builds, orchestration, and isolation security. Sandboxes sell that ops burden away. Compare long-lived boxes on our VPS provider compare; compare disposable ones here.
Not ranked — and why that's honest
- Morph — NOT CAPTURED: morph.so renders its pricing client-side — no rate constants in the served HTML or JS bundles, and its docs domain did not resolve from the capture network during the 2026-08-24 pass. Historical per-vCPU/RAM figures circulate in older comparisons, but we don't quote numbers we couldn't re-verify, so Morph renders unranked here.
- Fly Machines — NOT RANKED: Fly prices named CPU/RAM PRESETS as bundles (e.g. performance-2x/4GiB $0.0861/hr; shared-cpu-4x/8GiB $0.0594/hr) — the bundle does not decompose into a per-vCPU + per-GiB pair that holds across presets (implied RAM ranges ≈$7–8/GiB-hr depending on which rows you difference), so an honest fit into our two-axis model isn't possible without inventing numbers. Per-second billing confirmed; 40% reservation discounts exist (12-mo upfront commitments). Representative presets are quoted here so you can compare directly against ranked vendors' modeled outputs.
Who should skip what
- Northflank — Skip if you need agent-native session semantics (snapshot/pause-resume) — Northflank is general container platform pricing applied to long-running services and jobs, not a purpose-built agent sandbox SDK.
- Daytona — Skip if you want the most battle-tested SDK ecosystem — E2B has the larger community around code-interpreter sandboxes at identical rates.
- E2B — Skip if your agents need long-lived environments — Hobby sessions die at 1h and Pro at 24h; this is for short-lived code-running sandboxes.
- Cloudflare Containers — Skip if you want vendor-managed snapshot/resume semantics — Containers scale to zero but have no cross-machine session pause/resume story like E2B/Daytona snapshots.
- Beam (sandbox) — Skip if you want the biggest platform — young product; fine for browser/computer-use sandboxes, thinner on compliance and regions than E2B or Modal.
- Modal (Sandboxes) — Skip if pure sandbox price is the decision — you pay ~40% over E2B/Daytona for the same vCPU-hour; Modal earns it back only when the rest of your stack already lives on Modal functions.
- Self-host Firecracker (Hetzner) — Skip if usage is bursty or spiky — the host bills 24/7 whether sandboxes run or not; below roughly half utilization the managed providers win on total cost and you still own the pager.
- Runloop — Skip unless you specifically want its Devbox/Blueprint workflow — the highest per-hour rates captured here, with add-on fees on top.
Price-change register updated 2026-08-25
| Vendor | $/vCPU-hr | $/GiB-hr | Previous | Δ vCPU-hr | Captured |
|---|---|---|---|---|---|
| Northflank | $0.0167 | $0.0083 | — | first capture | 2026-08-24 |
| Daytona | $0.0504 | $0.0162 | — | first capture | 2026-08-24 |
| E2B | $0.0504 | $0.0162 | — | first capture | 2026-08-24 |
| Beam (sandbox) | $0.0675 | $0.0230 | — | first capture | 2026-08-24 |
| Modal (Sandboxes) | $0.0710 | $0.0240 | — | first capture | 2026-08-24 |
| Cloudflare Containers | $0.0720 | $0.0090 | — | first capture | 2026-08-24 |
| Runloop | $0.1080 | $0.0252 | — | first capture | 2026-08-24 |
| Self-host Firecracker (Hetzner) self-host · host $153.59/mo | — | — | — | first capture | 2026-08-24 |
Every row carries its own capture date from vendor pricing pages; Δ compares the current capture to the previous distinct one (“capture refined” means our recording improved, not that prices changed). Uncaptured rates render as “—”, never zero; all rates are USD list pricing, never converted. Machine-readable copy: register.csv.
Commitments & regional reality
Computed at the default profile (500 sandbox-hours × 2 vCPU × 4 GiB) from vendors' own published constants — reservation-block terms, regional multipliers, and the honest absence of commitments elsewhere.
| Vendor / scenario | On-demand /mo | Scenario /mo | Δ vs on-demand | Reality check |
|---|---|---|---|---|
| Fly Machines (performance-2x/4GiB) | $43.05 | $25.83 | -40.0% 12-mo reserved block, −40% | Published preset bundle (2 vCPU / 4 GiB) — matches the default compare profile. Block credits do NOT roll over: an under-used month is still billed. |
| Fly Machines (shared-cpu-4x/8GiB) | $29.70 | $17.82 | -40.0% 12-mo reserved block, −40% | Published preset bundle (4 vCPU / 8 GiB). Block credits do NOT roll over: an under-used month is still billed. |
| Modal (Sandboxes) | $119.00 | $208.25 | +75.0% worst-case region ×1.75 | Regional multipliers run 1.5–1.75× on the base card (excluded from the main ranking — surfaced here instead). Non-preemptible containers add another 3× on top of whatever region you land in. |
| E2B | $82.80 | — | — | No public commitment/reserved pricing published as of the capture date — on-demand list is the entire price book. Any discount here would be invented. |
| Daytona | $82.80 | — | — | No public commitment/reserved pricing published as of the capture date — on-demand list is the entire price book. Any discount here would be invented. |
| Beam (sandbox) | $113.58 | — | — | No public commitment/reserved pricing published as of the capture date — on-demand list is the entire price book. Any discount here would be invented. |
| Runloop | $158.40 | — | — | No public commitment/reserved pricing published as of the capture date — on-demand list is the entire price book. Any discount here would be invented. |
| Self-host Firecracker (Hetzner) | $153.59 | — | — | No public commitment/reserved pricing published as of the capture date — on-demand list is the entire price book. Any discount here would be invented. |
| Morph | — | — | — | No verifiable public rate card captured this pass — nothing to model commitments against. |
| Cloudflare Containers | $90.00 | — | — | No public commitment/reserved pricing published as of the capture date — on-demand list is the entire price book. Any discount here would be invented. |
| Northflank | $33.33 | — | — | No public commitment/reserved pricing published as of the capture date — on-demand list is the entire price book. Any discount here would be invented. |
Sandbox buying guides
Head-to-head
All C(10,2) = 45 pairings have their own page — browse them all on the comparisons hub, or start with these analyst-curated duels:
Frequently asked questions
- How much does E2B cost per month?
- E2B bills per second at $0.0504 per vCPU-hour plus $0.0162 per GiB-hour of RAM. At 500 sandbox-hours/month on the default 2 vCPU / 4 GiB profile that is about $82.80/month of usage, after a $100 one-time free credit.
- Is it cheaper to self-host code sandboxes with Firecracker?
- Only at sustained high utilization. A dedicated-capacity cloud host able to run about eight 2-vCPU/4-GiB sandboxes bills roughly $150/month around the clock whether sandboxes run or not — below roughly half utilization, per-second-billed managed providers are cheaper and you avoid running the host yourself.
- Why is Modal more expensive for sandboxes than its advertised CPU pricing?
- Modal publishes two rate cards: general-purpose CPU functions (about $0.0236/vCPU-hr) and its separate Sandboxes product (about $0.142/core-hr, i.e. $0.071/vCPU-hr plus $0.024/GiB-hr memory). Agent-sandbox workloads bill on the higher Sandbox card.
- Which vendor is cheapest for AI agent sandboxes?
- Among metered pay-as-you-go vendors, Northflank holds the computed price floor at every workload profile ($33.33/mo at the 500 hr × 2 vCPU / 4 GiB reference — 2.5× under the E2B/Daytona tie). At sustained volume (roughly 1,000 sandbox-hours/month and up on typical profiles) the fixed-cost self-host Firecracker baseline becomes the outright cheapest. Rank any workload yourself with the calculator above.
Methodology
- All rates are published list prices fetched from official pricing pages on 2026-08-25; every row links its source.
- Model: compute = sandbox-hours × vCPUs × $/vCPU-hr; RAM bills separately at every captured vendor (hours × GiB × $/GiB-hr). Per-second and per-millisecond billing is normalized to hourly rates.
- Self-host baseline: a Hetzner CPX62 (16 vCPU / 32 GiB, incl. one IPv4) divided across its concurrent capacity at the reference profile (8 × [2 vCPU / 4 GiB]); other profiles fit min(vCPU-bound, RAM-bound) sandboxes per host — capacity is limited by the tightest resource, not the vCPU×RAM product. It deliberately EXCLUDES ops labor, image pipelines, backups, and idle time — the costs that usually kill the self-host business case.
- Excluded line items are stated per vendor: storage beyond free tiers, snapshot/object-storage fees, Axon coordination fees (Runloop), regional multipliers and non-preemptible surcharges (Modal), Windows images and GPU tiers (Daytona).
- Morph renders unranked because its pricing page ships no machine-readable rates and we don't quote numbers we couldn't re-verify this pass.