# Cua Fleet

Official pricing: https://cua.ai/pricing  
Category: agent-sandbox · Isolation: vm

## Pricing regimes (raw)

- **Fleet pool, Ubuntu 24.04 desktop VM (per warm replica-hour)** (resource): $0.044625/vCPU-h, $0.0223125/GiB-h
- **Fleet pool, Windows Server 2022 desktop VM** (resource): $0.044625/vCPU-h, $0.0223125/GiB-h
- **Cloud macOS (Apple Silicon, Lume/VZ), price not published** (resource) [sales]

## Features

Yes: idle auto-stop, ≥24 h sessions, custom image (Docker or snapshot), full VM (own kernel), Docker inside, root, browser, desktop GUI, computer-use API, browser automation API, desktop control API, browser + desktop control, SSH, HTTPS preview URLs, self-hosting, BYOC, open source, SOC 2, arm64, Windows, macOS, Python, Node.js, MCP server, gVisor or VM (no shared kernel), inbound access rules, firewall inside (nftables)

No: snapshots, memory snapshots, fork/clone, pause/resume, persistent disk, volumes, start from your own snapshot, your own Docker/OCI image, anti-bot stealth, CAPTCHA solving, residential IPs, preinstalled agents, public IPv4, egress allowlist, static egress IP, HIPAA, SSO, GPU, GPU desktop, HTTP method/path egress rules, wake on request, live resize, live fork (no pause), memory fork, EU data residency, automatic snapshots, snapshots on demand, agent harness API, hosted agent API (their own agent), secret proxy, secret proxy for any API, extra volumes, shared volumes

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

## Caveats

- Only CPU and memory rates are published; storage, snapshots, egress, IPv4, billing granularity, minimums, concurrency and free credits are unpublished (null).
- Pool-based billing: replicas consume capacity while warm, independent of claims. The engine's session hours only match cost if the pool is scaled to demand (see warm_pool_overhead knob).
- Windows priced at the same rates as Linux per the single published card; a Windows licence surcharge is not listed (unverified).
- Homepage markets Linux, Windows, macOS and Android, but SDK 0.7.0 docs accept only Ubuntu 24.04 VM and Windows Server 2022 VM images for Fleet; macOS/Android/Windows 11 are local-only. Card sets macos false.
- Isolation for Fleet not stated; local runtimes are QEMU/Hyper-V VMs (card isolation 'vm').
- SOC 2 Type I (not Type II). BYOC and on-prem available via sales.
- Legacy 2025 Cua Cloud Sandbox sizes (Small 1/4, Medium 2/8, Large 8/32) had no published prices; superseded by Fleet per-resource pricing.
- Third-party guides mention a free dev tier; not on the official pricing page (free credits set 0).
- No benchmark data in benchmarks.json.
- Windows: Cua images do not include a commercial Windows licence (BYOL); the estimate excludes licence cost. Source: https://cua.ai/docs/reference/sandbox-sdk/os-image-catalog
- Snapshots: marketing says 'fork from snapshots', but cua-sandbox 0.7.0 docs say Sandbox.snapshot() is unimplemented for Fleet and Fleet rejects snapshot-derived images; card keeps snapshot 'none' per the versioned docs (conflict noted).
- Closing a connection, releasing a claim and deleting a pool are separate; warm pool capacity keeps billing until the pool is scaled down/deleted (keep_pool can preserve pools).
- 2026-09-28 re-check: SSH not documented for Cua Fleet (docs describe computer-server as the control channel), so features.ssh stays null. Disk size per Fleet VM and any disk/storage price are unpublished (cua.ai/pricing lists only $0.044625/vCPU-h and $0.0223125/GiB-h); included_disk_gib and disk_gib_month stay null (unknown, not free).
- Hosted macOS ("Cloud macOS") exists but has no published price; its mode is unpriced and flagged sales.
- Fleet VMs run on KubeVirt (KVM VMs orchestrated by Kubernetes); pool templates take cpu_cores and memory (e.g. 4 cores / 8Gi Linux, 4 cores / 4Gi Windows in the Terraform examples); no published max size.
- Custom Fleet images must be a prebuilt OCI containerDisk artifact carrying a bootable guest disk (/disk/disk.img); Dockerfiles / ordinary app containers are not accepted (image_oci false is correct).
- Homepage now markets 'Windows 11 and Server 2025', but the SDK 0.7.0 docs still map only Windows Server 2022 (and reject Windows 11) for Fleet; Windows images carry no licence (BYOL).

## How this provider charges


Cua is an open-source (MIT) computer-use sandbox framework plus a hosted service, **Cua Fleet**, that runs GUI
desktop VMs (Ubuntu 24.04 and Windows Server 2022 in the cloud today). The only published price is one per-resource
card: **$0.044625/vCPU-h** and **$0.0223125/GiB-h** (1 vCPU + 2 GiB = $0.08925/h). 4 vCPU / 8 GiB =
0.1785 + 0.1785 = **$0.357/h**. No plan fees, storage, egress, OS surcharge or free credits are published.

The key billing trap is structural: Fleet is **pool-based**. A pool keeps N warm replicas that "continue to consume
capacity until it is scaled down or deleted"; a claim merely borrows one warm computer. You pay for warm replica
uptime, not for claimed/active time.

## Regime table

| Regime | When it applies | How billed | Numbers | Source |
|---|---|---|---|---|
| Local open-source Cua Sandbox | Run on your own machine (Docker, QEMU, Apple VZ/Lume, Android emulator) | Free (MIT); you pay your own hardware | $0 | https://cua.ai/pricing, https://cua.ai/docs/reference/sandbox-sdk/runtime-support.md |
| Cua Fleet usage-based (Linux) | Hosted Ubuntu 24.04 VM pools | Allocated vCPU + memory per hour of replica uptime | $0.044625/vCPU-h + $0.0223125/GiB-h | https://cua.ai/pricing |
| Cua Fleet usage-based (Windows) | Hosted Windows Server 2022 VM pools | Same published rates; no Windows licence surcharge listed (unverified) | same | https://cua.ai/pricing, https://cua.ai/docs/reference/sandbox-sdk/runtime-support.md |
| Warm pool capacity | Pool created with `replicas=N` | Each warm replica bills from creation until scale-down/delete, **claimed or not**; "cloud resources can incur usage charges after your workload finishes" | N × size × rate | https://cua.ai/docs/how-to-guides/sandbox/create-fleet-capacity.md, https://cua.ai/docs/cloud-fleets.md |
| Claim | `claim` one computer from a pool | No separate claim fee documented; released on context exit | $0 extra (unpublished) | https://cua.ai/docs/cloud-fleets.md |
| TTL expiry | `ttl_seconds_after_created` on pools/claims | Deletes resource N s after **creation** (not idle); no default TTL = pool bills until deleted | 0–4,294,967,295 s | https://cua.ai/docs/how-to-guides/sandbox/expire-pools-and-claims, https://cua.ai/docs/reference/sandbox-sdk/pool.md |
| macOS / Android cloud | Marketed on homepage | **Not accepted by Fleet** in SDK 0.7.0 (local only) | n/a | https://cua.ai/, https://cua.ai/docs/reference/sandbox-sdk/runtime-support.md |
| BYOC / on-prem (Enterprise) | Run Fleet in your cloud or datacenter | Contact sales | null | https://cua.ai/pricing |
| Storage / snapshots / egress | — | Not published | null | https://cua.ai/pricing |
| Legacy Cua Cloud Sandbox (2025) | Launch 2025-05-28: Small 1 vCPU/4 GB, Medium 2/8, Large 8/32 | "pay only for compute time"; prices never published in the post; superseded by Fleet per-resource pricing | null | https://cua.ai/blog/introducing-cua-cloud-containers |

No dated price change with numbers found. Third-party guides mention a free development tier; not on the official page (unverified).

## Gotchas

1. **You pay for warm pool replicas, not claims.** A pool of 50 kept up 24/7 costs 50 × 730 h regardless of how many agent runs happen. Scale replicas to 0/delete the pool between shifts.
2. **TTL counts from creation, not last use**, and there is no default TTL — forgotten pools bill forever.
3. **Windows costs the same as Linux** on the published card; no licence surcharge is listed (confirm before relying on it).
4. **macOS and Android are marketing, not Fleet reality** in SDK 0.7.0 (local only).
5. RAM is priced at half the vCPU rate per GiB — memory-heavy desktops (1:4 ratios like the old Small 1/4 size) cost more per vCPU than E2B-style 1:2 shapes.
6. Storage, egress, IPv4, granularity, minimums, concurrency limits and free credits are all unpublished.

## Worked example

4 vCPU / 8 GiB, 50 concurrent × 8 h/day × 22 days, 30% CPU, 50 GiB snapshots, 100 GiB egress.

| Regime | Feasible? | Monthly total |
|---|---|---|
| Fleet pool scaled to 50 replicas only during shifts (8,800 replica-h) | Yes (limits unpublished) | 8,800 × $0.357 = **$3,141.60** + unknown storage/egress |
| Fleet pool of 50 left warm 24/7 | Yes | 50 × 730 × $0.357 = **$13,030.50** |
| Fleet on Windows Server 2022 | Yes | same as Linux at published rates ($3,141.60) |
| BYOC / on-prem | Enterprise | unpublished Cua fee + own infrastructure |
| Local OSS | Your hardware | $0 to Cua |

30% CPU does not matter (allocation billing). Snapshot and egress costs unknown (null).