# Sail Research (Sailboxes)

Official pricing: https://docs.sailresearch.com/pricing  
Category: agent-sandbox · Isolation: firecracker

## Pricing regimes (raw)

- **Sailbox, billed on observed usage (used vCPU / used RAM / used disk)** (resource): $0.015/vCPU-h, $0.008/GiB-h

## Features

Yes: snapshots, memory snapshots, fork/clone, pause/resume, persistent disk, volumes, idle auto-stop, ≥24 h sessions, custom image (Docker or snapshot), start from your own snapshot, your own Docker/OCI image, full VM (own kernel), Docker inside, root, systemd, preinstalled agents, SSH, HTTPS preview URLs, custom domains, raw TCP inbound, egress allowlist, open internet, SOC 2, HIPAA, arm64, Python, Node.js, credential injection, HTTP method/path egress rules, wake on request, live fork (no pause), memory fork, scheduled runs (cron), zero data retention, automatic snapshots, snapshots on demand, gVisor or VM (no shared kernel), inbound access rules, firewall inside (nftables), secret proxy, secret proxy for any API, extra volumes, shared volumes

No: nested virtualization, browser, desktop GUI, computer-use API, code interpreter, browser + desktop control, anti-bot stealth, CAPTCHA solving, residential IPs, public IPv4, static egress IP, self-hosting, BYOC, open source, SSO, GPU, Windows, macOS, GPU desktop, live resize, webhooks, EU data residency, MCP server, agent harness API, hosted agent API (their own agent), fast boot (benchmarked)

Unknown: everything else. Evidence (source + quote) per feature: https://battleships.dev/data/providers/sail.json → feature_evidence

## Caveats

- Usage basis cited (docs.sailresearch.com/sailboxes-pricing, re-checked 2026-09-28): "Used vCPU/hour" = "Guest /proc/stat CPU time, sampled about every 15 s" ($0.015); "Used RAM (GiB)/hour" = "Guest MemTotal minus MemAvailable, sampled about every 15 s" ($0.008); "Usage accrues only while a Sailbox is running." So ram_basis active is the documented meter, not an assumption.
- Both CPU and RAM are billed on observed usage, not allocation: engine uses cpu_util and ram_util. RAM 'used' = MemTotal - MemAvailable, so reclaimable page cache is not billed; ram_util should be the working set, not the allocation.
- Used NVMe disk ($0.0007/GiB-h = $0.511/GiB-month at 730 h) accrues only while running; the engine charges disk_gib_month per month, which over-counts disk for part-time sandboxes (true cost = GiB x running hours x $0.0007). Disk is 'used bytes on /' so the image footprint counts.
- Sleeping/paused state (memory + disk checkpoint) is free per docs, hence snapshot_gib_month 0. Storage price for explicit checkpoints (default 7-day TTL) is not published. Volumes cost $0.000411/GiB-h (~$0.30/GiB-month) for every hour they exist, even while Sailboxes sleep (not modelled).
- Creation fee depends on size (S .005 / M .01 / L .012) and applies on Free only (Pro/Enterprise set waives_start_fee); start_fee uses the M fee.
- Granularity: usage sampled ~every 15 s; exact rounding not documented. No minimum billed time published.
- Credits are prepaid and org-wide, shared with Sail's LLM inference API. Pro's included $100/month is not a fee credit (fee_is_credit false); first month adds $150.
- Free plan's $5/month requires a payment method on file (free.monthly_credit = Free plan included_usd; don't subtract both).
- Egress/ingress pricing not published (null). No public IPv4; published ports serve HTTP or raw TCP (with source allowlist).
- Regions for Sailboxes not documented (DPA says service providers in multiple regions). Isolation: docs say 'VM' / live-migrated VMs; hypervisor not disclosed (providers/sail.json says firecracker, unverified).
- Sailboxes are live-migrated a few times a day for bin-packing (seconds-long latency blip); any wake can come back cold (disk intact, processes gone).
- Enterprise volume pricing unpublished. The ComputeSDK 'Sail' provider is Sail Research (sailresearch.com), not sail.dev (Coder).
- Hypervisor confirmed: Sail's own engineering blog (2026-07-14) says every Sailbox is a Firecracker microVM with virtio-mem memory hot-plug; supersedes the older caveat that the hypervisor is undisclosed. The fleet runs on ephemeral compute with speculative background checkpoints.
- New organizations are capped on requests in flight for their first ten minutes (429 with Retry-After); no published number. No per-minute Sailbox creation rate limit is published beyond the plan concurrency caps (100 Free / 5,000 Pro).
- Checkpoint copies: 'Starting several copies from one checkpoint reuses the same checkpoint data, so only the first copy pays for it' - implies some checkpoint-data cost on the first copy that is not itemised on the pricing page; checkpoint storage price remains unpublished.

## How this provider charges


Sail Research sells two products: an LLM inference API and Sailboxes, which are persistent Linux VMs for long-horizon
agents. Both draw on one prepaid credit balance. Sailboxes are the only sandbox in this set that bills **every
dimension on observed usage**: used vCPU, used RAM and used disk, sampled about every 15 s. The chosen size (s/m/l)
only sets a ceiling. A Sailbox that is sleeping, paused, checkpointing or cold-starting costs nothing, and autosleep
kicks in after **30 s** of full idleness by default. The ComputeSDK "Sail" provider is Sail Research
(sailresearch.com), not sail.dev, which redirects to Coder.

Base rates: used vCPU **$0.015/h**, used RAM **$0.008/GiB-h**, used NVMe disk **$0.0007/GiB-h** (while running only),
volumes $0.000411/GiB-h (always). At 100% of 4 vCPU with 8 GiB resident: 0.06 + 0.064 = **$0.124/h**, about 37% of E2B's allocation price.

## Regime table

| Regime | When it applies | How billed | Numbers | Source |
|---|---|---|---|---|
| Free (PAYG, prepaid credits) | Default | No fee. Prepaid credits. $5/month free credit once a card is attached. Creation fee per Sailbox and per copy | $0 fee; $5/mo credit; 4 seats; 100 concurrent; creation S $0.005 / M $0.01 / L $0.012 | https://docs.sailresearch.com/pricing |
| Pro | Need >100 concurrent or >4 seats, or want creation fees waived | $250/mo **pure fee** plus $100/mo included credits (+$150 in the first month). Usage beyond that draws on prepaid credits | $250; $100/mo credit; 5,000 concurrent; unlimited seats; creation fees waived | https://docs.sailresearch.com/pricing |
| Enterprise | Contract | Volume pricing billed monthly in arrears; HIPAA BAA, SLAs | unpublished | https://docs.sailresearch.com/pricing |
| Running (usage-based) | Sailbox awake | CPU = guest /proc/stat CPU time; RAM = MemTotal − MemAvailable (page cache excluded); disk = statfs used bytes on `/`; sampled ~every 15 s. Size is a ceiling, not a reservation | $0.015 vCPU-h, $0.008 GiB-h, $0.0007 GiB-h disk | pricing-page tooltips; https://docs.sailresearch.com/sailboxes-billing |
| Sizes | Chosen at create (default `m`) | Set the vCPU count, the RAM/disk ceilings and the creation fee. No idle cost | s 1 vCPU / 16 GiB (2–64) / 32 GiB (8–128); m 4 / 32 (8–128) / 128 (32–512); l 8 / 64 (16–256) / 256 (64–1024) | https://docs.sailresearch.com/sailboxes-billing, CLI reference |
| Autosleep | Fully idle (no CPU, no timers, no open connections) for the idle window | Memory+disk checkpoint. **$0** for compute and disk. Wakes on traffic, exec, SSH, a reply or an alarm | default 30 s; configurable 1–3600 s or `never` | https://docs.sailresearch.com/sailboxes-autosleep |
| Sleep while awaiting an LLM reply | Process waits on an outbound TCP/HTTP reply with timeout ≥ 5 min | Sleeps; the reply wakes it (force-wake after ~5 min) | $0 while asleep | sailboxes-autosleep |
| Paused | `pause` (only an explicit `resume` wakes it) | Not billed | $0 | sailboxes-billing FAQ |
| Checkpoints / copies | `checkpoint`, start-from-checkpoint | The checkpoint handle expires after 7 days by default (TTL configurable). **Storage price not published.** Each copy bills its own usage plus a creation fee on Free | null | SDK reference; sailboxes-billing FAQ |
| Volumes | Mounted shared storage | Billed every hour the volume exists, including while Sailboxes sleep | $0.000411/GiB-h ≈ $0.30/GiB-mo | https://docs.sailresearch.com/pricing |
| Credits exhausted | Balance hits 0 | All running and sleeping Sailboxes are **paused**. New spend (create, resume, exec) is blocked. Nothing is terminated | — | sailboxes-billing FAQ |
| Terminated | `terminate` or `max_lifetime_seconds` | Billing stops. Anything not on a volume or checkpoint is lost | $0 | API reference |
| Egress / IPv4 | — | Not published. No public IPv4; published ports serve HTTP or raw TCP | null | — |
| Inference surcharges (not sandbox) | Pro/Enterprise US-only inference | +20% (Pro) / +10% (Enterprise) on token prices | n/a to Sailboxes | pricing page |

No dated Sailbox price changes found.

## Gotchas

1. **RAM is billed on what the guest actually uses** (MemTotal − MemAvailable), not on the 32 GiB ceiling. A 30%-busy agent holding 2 GiB pays for 2 GiB. Estimators that use allocated RAM overstate Sail by a wide margin.
2. **Autosleep is aggressive (30 s) and on by default.** It even sleeps while an agent waits on a long LLM HTTP call, so wall-clock session hours overstate billed hours. The flip side: any wake can come back **cold** (disk intact, processes gone), and a request timeout under 5 min keeps the box awake.
3. **Disk is "used bytes on /" and bills only while running**, so the image footprint counts but sleeping disk is free. Volumes are the exception and bill 24/7.
4. **The Pro $250 is a fee, not credit.** Only $100/mo comes back as credit (first month $250). At ≤ 100 concurrent, Free is cheaper unless you need seats or 5,000 concurrency.
5. **Creation fees apply to every copy/fork on Free.** They are tiny ($0.005–0.012), but a fan-out of 10k copies costs $50–120.
6. **Credits are prepaid and shared with inference.** Heavy token usage can drain the balance, and when it hits zero every Sailbox is paused.
7. **Live migration** happens a few times a day (seconds-long blip).
8. Checkpoint storage, egress and regions are unpublished.

## Worked example

Workload: 4 vCPU / 8 GiB (size `m` with `memory_limit_gib=8`), 50 concurrent × 8 h/day × 22 days = **8,800 Sailbox-hours**,
30% CPU, 50 GiB snapshots, 100 GiB egress. Sailboxes are kept and slept between shifts (50 creations). Assumed 10 GiB used disk.

| Line | Calculation | $ |
|---|---|---|
| CPU (used) | 4 × 0.30 × 0.015 × 8,800 | 158.40 |
| RAM, 8 GiB resident all session | 8 × 0.008 × 8,800 | 563.20 |
| RAM, if only 4 GiB resident | 4 × 0.008 × 8,800 | 281.60 |
| Used disk 10 GiB (running hours only) | 10 × 0.0007 × 8,800 | 61.60 |
| Snapshots / sleep state 50 GiB | sleeping not billed; explicit checkpoint price unpublished | 0 (null) |
| Egress 100 GiB | unpublished | null |
| Creation fees (Free) | 50 × $0.01 (fresh box per session: 1,100 × $0.01 = $11) | 0.50 |

| Regime | Feasible? | Monthly total |
|---|---|---|
| Free, 8 GiB resident | Yes (50 ≤ 100 concurrent) | **$778.70** = 783.20 + 0.50 − 5 |
| Free, 4 GiB resident | Yes | **$497.10** |
| Free, awake 50% of the session (autosleep during LLM waits; CPU-seconds unchanged) | Yes | **$466.30** = 158.40 + 281.60 + 30.80 + 0.50 − 5 |
| Pro, 8 GiB resident | Yes | **$933.20** = 250 + 783.20 − 100 (first month $783.20) |
| Enterprise | Yes | unpublished volume pricing |
| Anti-pattern: autosleep `never`, left up 24 h/day (8 GiB resident, ~0 CPU idle) | Yes | ≈ **$2,028** on Free (17,600 extra idle h × $0.071) |

For comparison, E2B Pro on the same workload costs $3,064.56.