# Cloudflare Sandbox SDK / Containers

Official pricing: https://developers.cloudflare.com/containers/platform/pricing/  
Category: agent-sandbox · Isolation: vm

## Pricing regimes (raw)

- **Pay-as-you-go (active CPU, provisioned RAM/disk)** (resource): $0.072/vCPU-h, $0.009/GiB-h

## Features

Yes: snapshots, 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, code interpreter, preinstalled agents, SSH, HTTPS preview URLs, custom domains, egress allowlist, open internet, open source, SOC 2, HIPAA, SSO, Python, Node.js, credential injection, HTTP method/path egress rules, wake on request, audit logs, EU data residency, snapshots on demand, agent harness API, fast boot (benchmarked), gVisor or VM (no shared kernel), inbound access rules, secret proxy, secret proxy for any API, extra volumes, shared volumes

No: memory snapshots, fork/clone, pause/resume, persistent disk, nested virtualization, browser, desktop GUI, computer-use API, browser + desktop control, anti-bot stealth, CAPTCHA solving, residential IPs, public IPv4, raw TCP inbound, static egress IP, private networking, self-hosting, BYOC, GPU, arm64, Windows, macOS, GPU desktop, live resize, live fork (no pause), memory fork, webhooks, spend limits, MCP server, automatic snapshots, hosted agent API (their own agent), firewall inside (nftables)

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

## Caveats

- 4 vCPU / 8 GiB is not offered: >=3 GiB RAM per vCPU, so 4 vCPU always pays 12 GiB; max shape 4 vCPU/12 GiB/20 GB.
- Default sleepAfter is 10 min: RAM+disk bill at full size during the idle tail after each activity burst (~$0.113/h for standard-4); keepAlive bills continuously with no cap.
- Sleep wipes all state (no paused state). Persistence = directory backups to your own R2 bucket, billed at R2 Standard $0.015/GB-month after 10 GB-month free (+ Class A $4.50/M, Class B $0.36/M after free tiers); backup TTL does not delete R2 objects. Whole-container snapshots are private beta, unpriced.
- Each sandbox is a Durable Object: $0.15/M requests after 1M, $12.50/M GB-s after 400k GB-s at 128 MB; if awake for the whole container run ~ $0.00576/h per sandbox (unverified). Worker requests $0.30/M after 10M; Workers Logs $0.60/M after 20M. Third-party estimate: +5-10%.
- Egress free allotment/rate vary by region: NA/EU 1 TB then $0.025/GB; Oceania/Korea/Taiwan 500 GB then $0.05/GB; elsewhere 500 GB then $0.04/GB. Card uses NA/EU.
- Instance type is per container class at deploy time, not per sandbox.
- No guaranteed runtime: irregular host restarts stop instances (SIGTERM, 15 min grace) and wipe state.
- Concurrency limited by account resource caps (1,500 vCPU / 6 TiB / 30 TB), not a sandbox count.
- perf.cpu_runs_s: no hpc benchmark for Cloudflare; dax never succeeded in ComputeSDK runs.
- open_source refers to the SDK/container server/bridge only; platform is proprietary and cannot be self-hosted.
- Verifier 2026-09-28: storage.included_disk_gib was 20 (engine treated 20 GB as free). Cloudflare bills provisioned disk at $0.00000007/GB-s while the container runs (https://developers.cloudflare.com/containers/pricing/), so included_disk_gib -> 0 and disk_gib_month -> 0.18396 (= $0.000252/GB-h x 730); with disk_billed_when_stopped=false the engine bills it only for running hours, which matches. The 200 GB-h/month included disk sits inside Workers Paid included_usd. Max disk is 20 GB (<= 2 GB per GiB RAM), which the engine does not enforce.
- Verifier 2026-09-28: snapshot_free_gib_account = 10 added (R2 Standard 10 GB-month free, https://developers.cloudflare.com/r2/pricing/).
- Account concurrency caps (1,500 vCPU / 6 TiB / 30 TB) and per-sandbox Durable Object / Worker request charges are not modelled by the engine.
- Per container class, wrangler max_instances defaults to 20 concurrently running instances; a start request beyond it errors. Self-serve config change up to the account caps (1,500 vCPU / 6 TiB / 30 TB).
- Durable Object usage above the included amount is rounded up to the next billable unit (1M GB-s / 1M requests) before the rate applies, so the first DO overage is at least $12.50 (duration) or $0.15 (requests).
- root=true means uid 0 inside an unprivileged container: no privileged mode, Docker only via docker:dind-rootless with iptables disabled, no kernel modules.
- Worker subrequests are not billed; each Sandbox SDK operation is an RPC call on the sandbox's Durable Object stub and is billed as one DO request ($0.15/M after the 1M/month included), not as a Worker request.
- DO duration may not run for the whole container lifetime: DOs idle and eligible for hibernation are not billed for duration, and an inactivity timeout can keep the container running while the DO sleeps. The ~$0.00576/h per-sandbox DO estimate is an upper bound.

## How this provider charges


As of 2026-09-28. Sandbox SDK has no price list of its own: it is billed as **Containers** (vCPU, memory, disk, egress) **plus** the Worker and Durable Object behind each sandbox, **plus** Workers Logs if enabled, **plus** R2 for any backups or bucket mounts.

Core rates (Workers Paid, standard, not Enterprise):

| Resource | Rate | Hourly equivalent | Basis | Included / month |
|---|---|---|---|---|
| CPU | $0.000020 / vCPU-s | $0.072 / vCPU-h | **active usage only** (since 2025-11-21) | 375 vCPU-min (6.25 vCPU-h, $0.45) |
| Memory | $0.0000025 / GiB-s | $0.009 / GiB-h | **provisioned** for the instance type | 25 GiB-h ($0.225) |
| Disk | $0.00000007 / GB-s | $0.000252 / GB-h | **provisioned** for the instance type | 200 GB-h ($0.05) |
| Granularity | 10 ms | | billed "while actively running": from the first request or manual start until the instance sleeps | |

## Regime table

| # | Regime | When it applies | How it's billed | Numbers | Source |
|---|---|---|---|---|---|
| 1 | **Workers Free plan** | Account with no subscription | Containers/Sandbox are **not available** ("N/A") | $0 but unusable | https://developers.cloudflare.com/containers/pricing/ |
| 2 | **Workers Paid plan** (base) | Any account that uses Sandbox | $5/month account minimum. It is **not a usage credit**: it covers the included allotments only | $5/mo. Includes 375 vCPU-min, 25 GiB-h, 200 GB-h of Containers (about $0.73 of container usage), 10M Worker requests, 30M Worker CPU-ms, 1M DO requests, 400k DO GB-s | https://developers.cloudflare.com/workers/platform/pricing/ |
| 3 | **Enterprise contract** | Workers Enterprise accounts | Billed per contract ("reach out to your account team") | Not published (null) | https://developers.cloudflare.com/workers/platform/pricing/ |
| 4 | **Active running, active-CPU billing** (the normal regime) | Container is awake | CPU × actual utilisation, and RAM + disk × provisioned size, per 10 ms | $0.072 per active vCPU-h, $0.009/GiB-h, $0.000252/GB-h | https://developers.cloudflare.com/containers/pricing/ , https://developers.cloudflare.com/changelog/post/2025-11-21-new-cpu-pricing/ |
| 5 | **Awake but idle** (waiting for `sleepAfter`) | After the last request until `sleepAfter` expires (default **10 min**) | CPU is about $0 because it is active-only. **RAM and disk keep billing at full provisioned size** | standard-4 idle costs 12×0.009 + 20×0.000252 = **$0.113/h** | https://developers.cloudflare.com/containers/faq/ , https://developers.cloudflare.com/sandbox/concepts/sandboxes/ |
| 6 | **keepAlive: true** (always-on) | SDK option. Sends heartbeat pings every 30 s and never sleeps | Same as regime 4/5, continuously, until `destroy()`. There is **no monthly cap or commit discount** | standard-4 at 0% CPU: $0.113/h = $82.5 per 730 h. At 100% CPU: $0.401/h = $292.8 per 730 h | https://developers.cloudflare.com/sandbox/concepts/sandboxes/ |
| 7 | **Sleeping / stopped** | After `sleepAfter`, or on a host restart | $0 compute, $0 disk. **All state is wiped** (files, processes, sessions): there is no paused state to pay for | $0 | https://developers.cloudflare.com/containers/faq/ |
| 8 | **Cold start / boot** | Every wake from sleep | Billing starts "when a request is sent to the container or when it is manually started", so image pull and boot are billed at RAM+disk rates | Cold start "1-3 s" (docs). ComputeSDK median 2.0 s | https://developers.cloudflare.com/containers/pricing/ , https://developers.cloudflare.com/containers/faq/ |
| 9 | **Preset instance types** | `instance_type` in wrangler config (per container class, set at deploy, not per sandbox) | Same per-resource rates. Preset fixes the provisioned RAM and disk | lite 1/16 vCPU/256 MiB/2 GB; basic 1/4/1 GiB/4 GB; standard-1 1/2/4 GiB/8 GB; standard-2 1/6 GiB/12 GB; standard-3 2/8 GiB/16 GB; standard-4 4/12 GiB/20 GB. Hourly at 100% CPU: 0.00725 / 0.0280 / 0.0740 / 0.1290 / 0.2200 / 0.4010 | https://developers.cloudflare.com/containers/platform-details/limits/ |
| 10 | **Custom instance types** (GA 2026-01-05) | Custom vCPU/RAM/disk | Same rates. **1-4 vCPU, at least 3 GiB RAM per vCPU, at most 12 GiB, at most 20 GB disk, at most 2 GB disk per GiB RAM** | 4 vCPU/8 GiB is **impossible**: 4 vCPU forces 12 GiB | https://developers.cloudflare.com/containers/platform-details/limits/ , https://developers.cloudflare.com/changelog/product/containers/ |
| 11 | **Legacy provisioned-CPU billing** (historical) | Before 2025-11-21 | CPU billed on allocated vCPU × running time | Same $0.00002/vCPU-s, on allocation | https://developers.cloudflare.com/changelog/post/2025-11-21-new-cpu-pricing/ |
| 12 | **Egress: North America & Europe** | Traffic egressing in NA/EU | Account-wide allotment, then per GB | 1 TB/month free, then $0.025/GB | https://developers.cloudflare.com/containers/pricing/ |
| 13 | **Egress: Oceania, Korea, Taiwan** | Traffic egressing there | | 500 GB free, then **$0.05/GB** | same |
| 14 | **Egress: everywhere else** | Other regions (SAM, ME, AFR, rest of APAC) | | 500 GB free, then $0.04/GB | same |
| 15 | **Ingress** | Inbound | Free (no ingress price listed) | $0 | same |
| 16 | **Durable Object per sandbox** | Always: each sandbox is one DO | Requests: $0.15/M after 1M. Duration: $12.50 per M GB-s after 400k GB-s, billed on a 128 MB allocation "while the object actively runs or idles without hibernation eligibility" | If the DO is awake for the whole container run: 0.128 GB × 3600 s = 460.8 GB-s/h, about **$0.00576/h per sandbox** (unverified that the DO is billed for the whole run) | https://developers.cloudflare.com/durable-objects/platform/pricing/ , https://developers.cloudflare.com/sandbox/platform/pricing/ |
| 17 | **Worker requests / CPU** | Every SDK call routed through your Worker | $0.30/M requests after 10M. $0.02/M CPU-ms after 30M | The HTTP transport makes **each SDK op a subrequest** (1,000 per request cap on Paid). The RPC transport multiplexes them | https://developers.cloudflare.com/workers/platform/pricing/ , https://developers.cloudflare.com/sandbox/platform/limits/ |
| 18 | **Workers Logs** (optional) | Observability enabled | 20M log events/month included, then $0.60/M | | https://developers.cloudflare.com/workers/platform/pricing/ |
| 19 | **Directory backups** (GA 2026-02-23) | `createBackup(dir)` writes squashfs to **your own R2 bucket** | R2 Standard storage + ops. Backup TTL default 3 days, but expired objects are **not deleted** (you need R2 lifecycle rules) | $0.015/GB-month (10 GB-month free), Class A $4.50/M (1M free), Class B $0.36/M (10M free), R2 egress free. Infrequent Access: $0.01/GB-month, 30-day minimum, $0.01/GB retrieval | https://developers.cloudflare.com/r2/pricing/ , https://developers.cloudflare.com/sandbox/concepts/backup-restore/ |
| 20 | **Whole-container snapshots** | **Private beta** (contact your rep). Taken on idle before stop, best-effort | Pricing not published | null | https://developers.cloudflare.com/sandbox/tutorials/openai-agents-api/ |
| 21 | **Bucket mounts** (FUSE, R2/S3/GCS) | `mountBucket()` for persistence | Billed by the bucket provider (R2 rates above; R2 egress $0) | | https://developers.cloudflare.com/sandbox/guides/mount-buckets/ |
| 22 | **Image storage** | Registry images | 50 GB total image storage per account (a limit). No separate price found | null | https://developers.cloudflare.com/containers/platform-details/limits/ |
| 23 | **Account concurrency caps** | Running instances | 1,500 vCPU / 6 TiB RAM / 30 TB disk concurrent (raised 15x on 2026-02-25), raisable | For standard-4: at most 375 concurrent (vCPU-bound) | https://developers.cloudflare.com/changelog/product/containers/ |
| 24 | **Placement / jurisdiction** | `regions` / `jurisdiction: eu | fedramp` constraints (2026-04-05) | **No price multiplier** published for pinning. Only egress varies by region | x1 | https://developers.cloudflare.com/changelog/product/containers/ |
| 25 | **Warm pool** (self-hosted bridge) | Pre-started containers for instant boot | Each pooled container bills RAM+disk (regime 5) while it waits | pool size × $0.113/h (standard-4) | features/cloudflare-sandbox.json (bridge docs) |

## Gotchas

1. **You cannot buy 4 vCPU with 8 GiB.** The shapes are memory-heavy (at least 3 GiB per vCPU). A 4-vCPU sandbox always pays for 12 GiB (and up to 20 GB disk). RAM makes up more than half the bill at typical agent utilisation.
2. **Active-CPU applies to CPU only.** RAM and disk bill at full provisioned size from the first request until sleep, including the default **10-minute idle tail** after every burst of activity. For 5-minute sessions, that tail triples the RAM bill.
3. **The CPU rate is the most expensive per active vCPU-hour in the set** ($0.072, versus E2B at about $0.050 on allocation). Against E2B's 4 vCPU/8 GiB ($0.3312/h), a standard-4 ($0.288·util + $0.113/h) only wins below about 75% utilisation (verifier-corrected from "60-70%", 2026-09-28). At 100% util a standard-4 costs $0.401/h.
4. **Sleep = wipe.** There is no paused state, so nothing is billed while asleep, but persistence means R2 backups plus a restore on every wake (billed boot time plus R2 ops). Whole-container snapshots are private beta and unpriced.
5. **The $5 fee is not a credit.** It only includes about $0.73 of container resources. Everything above that is pay-as-you-go.
6. **Hidden per-sandbox surcharges.** One DO per sandbox (duration may run for the whole container lifetime, about $0.0058/h each), Worker requests (HTTP transport = one subrequest per SDK op), and Logs. A third party (bex.co) estimates **5-10% on top**.
7. **Backup TTL does not delete.** Expired backups stay in R2 and keep billing until a lifecycle rule removes them.
8. **Instance type is per container class (deploy-time), not per sandbox.** Mixed sizes need multiple classes.
9. **No guaranteed runtime.** Host restarts at irregular cadence stop instances (SIGTERM, then up to 15 min, then SIGKILL) and wipe state. An always-on keepAlive box is not truly durable.
10. **Egress doubles outside NA/EU** in some regions (verifier-corrected from "triples": $0.05 vs $0.025) ($0.05/GB in Oceania/Korea/Taiwan), and the free allotment halves (500 GB).
11. **Concurrency is resource-capped, not count-capped.** 1,500 vCPU means at most 375 standard-4 sandboxes at once, before a raise.

## Worked example

Workload: 4 vCPU / 8 GiB, 50 concurrent sandboxes × 8 h/day × 22 days = **8,800 sandbox-hours**, 30% CPU util, 50 GiB snapshots retained, 100 GiB egress.

Instance chosen: **standard-4 (4 vCPU / 12 GiB / 20 GB)** is the only fit, because custom 4 vCPU also forces 12 GiB or more. Included allotments are subtracted: 6.25 vCPU-h, 25 GiB-h, 200 GB-h. Snapshots are R2 Standard backups: 50 − 10 free = 40 GB × $0.015. Class A ops (about 1,100 backups) fall inside the 1M free. Egress of 100 GiB is inside the free allotment in every region. Assumes one wake per sandbox-day and `sleepAfter` idle tails excluded unless stated.

| Line | A: Paid, active CPU 30% (NA/EU) | B: same at 100% CPU (sustained build) | C: legacy provisioned-CPU (pre-2025-11-21, historical) | D: standard-3 downsize (2 vCPU/8 GiB, same absolute CPU work; does NOT meet 4 vCPU spec) |
|---|---|---|---|---|
| Plan fee | $5.00 | $5.00 | $5.00 | $5.00 |
| CPU | 10,560 − 6.25 = 10,553.75 vCPU-h × 0.072 = **$759.87** | 35,193.75 × 0.072 = $2,533.95 | $2,533.95 | 10,553.75 × 0.072 = $759.87 (60% util of 2 vCPU) |
| Memory | 105,600 − 25 = 105,575 GiB-h × 0.009 = **$950.18** | $950.18 | $950.18 | 70,375 × 0.009 = $633.38 |
| Disk | 176,000 − 200 = 175,800 GB-h × 0.000252 = **$44.30** | $44.30 | $44.30 | 140,600 × 0.000252 = $35.43 |
| Snapshots (R2) | $0.60 | $0.60 | $0.60 | $0.60 |
| Egress | $0 | $0 | $0 | $0 |
| **Subtotal (published rates)** | **$1,759.95** | **$3,534.03** | $3,534.03 | $1,434.28 |
| DO duration, if billed for the whole run (8,800 h × 460.8 GB-s − 400k = 3,655,040 GB-s × $12.50/M) | +$45.69 | +$45.69 | +$45.69 | +$45.69 |
| **Total incl. DO estimate** | **≈ $1,805.64** | ≈ $3,579.72 | ≈ $3,579.72 | ≈ $1,479.97 |

Other regime deltas on top of A:
- **Default 10-min idle tail** (1,100 sessions × 1/6 h × $0.11304/h of RAM+disk): **+$20.72**. With 5-minute agent tasks instead of 8 h sessions, this line would dominate.
- **keepAlive always-on** (50 × 730 h = 36,500 h at 30% CPU): CPU 43,793.75 × 0.072 = $3,153.15, RAM 437,975 × 0.009 = $3,941.78, disk 729,800 × 0.000252 = $183.91, plus fee → **$7,283.84** (+ DO (16,819,200 − 400k) GB-s × $12.50/M ≈ $205.24). There is no always-on discount.
- **Egress region**: 100 GiB is free everywhere. Above the allotment: $0.025 (NA/EU) / $0.04 (other) / $0.05 (OC/KR/TW) per GB.
- **Enterprise**: not published.
- Worker request, Worker CPU and Logs charges are workload-dependent and not included (small unless SDK ops are in the tens of millions).

Sources: https://developers.cloudflare.com/containers/pricing/ · https://developers.cloudflare.com/sandbox/platform/pricing/ · https://developers.cloudflare.com/containers/platform-details/limits/ · https://developers.cloudflare.com/changelog/post/2025-11-21-new-cpu-pricing/ · https://developers.cloudflare.com/changelog/product/containers/ · https://developers.cloudflare.com/containers/faq/ · https://developers.cloudflare.com/sandbox/concepts/sandboxes/ · https://developers.cloudflare.com/durable-objects/platform/pricing/ · https://developers.cloudflare.com/workers/platform/pricing/ · https://developers.cloudflare.com/r2/pricing/ · https://developers.cloudflare.com/sandbox/tutorials/openai-agents-api/ · https://bex.co/blog/2026/09/09/cloudflare-containers-sandboxes-active-cpu-pricing