# OpenComputer

Official pricing: https://opencomputer.dev/sandboxes  
Category: agent-sandbox · Isolation: vm

## Pricing regimes (raw)

- **Machine time per second (2 GB/1 vCPU default, auto-burst to 4/2 or 8/4)** (sizes): 0.25 vCPU / 0.5 GB (bursts to 1/2) 0.25 vCPU/0.5 GiB $0.0474/h; 0.5 vCPU / 1 GB (bursts to 2/4) 0.5 vCPU/1 GiB $0.0948/h; default 1 vCPU / 2 GB (bursts to 4/8) 1 vCPU/2 GiB $0.189/h; 2 vCPU / 4 GB (bursts to 8/16) 2 vCPU/4 GiB $0.378/h; 4 vCPU / 8 GB (bursts to 16/32) 4 vCPU/8 GiB $0.756/h [burstable]

## 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, full VM (own kernel), Docker inside, browser, browser automation API, anti-bot stealth, HTTPS preview URLs, egress allowlist, open internet, self-hosting, BYOC, open source, SSO, Python, Node.js, credential injection, HTTP method/path egress rules, wake on request, live resize, live fork (no pause), memory fork, webhooks, scheduled runs (cron), audit logs, MCP server, automatic snapshots, snapshots on demand, agent harness API, hosted agent API (their own agent), gVisor or VM (no shared kernel), inbound access rules, firewall inside (nftables), secret proxy, secret proxy for any API, shared volumes

No: your own Docker/OCI image, nested virtualization, desktop GUI, computer-use API, desktop control API, browser + desktop control, CAPTCHA solving, residential IPs, preinstalled agents, SSH, public IPv4, static egress IP, SOC 2, HIPAA, GPU, arm64, Windows, macOS, GPU desktop, fast boot (benchmarked), extra volumes

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

## Caveats

- Bare-sandbox pricing is now published separately (https://opencomputer.dev/sandboxes): machine time only, same per-unit rate as agent sessions, and 'Plans and prepaid credit on the agents page apply here too.'
- Dated change: July 2026 guides (2026-07-20) quoted '$0.004/min flat = $0.24/hr' for a 4 GB / 1 vCPU default with RAM resizable 1-16 GB; current pricing is $0.00315/min for 2 GB / 1 vCPU with bursts to 8 GB / 4 vCPU max.
- Pro/Max: the credit is 10x the fee and covers both tokens and machine time; encoded as fee + included_usd (not fee_is_credit). Rollover/expiry of unused credit not stated.
- Agent computers have a one-hour maximum running time and are replaced between commands (PR #806, 2026-09-25); sessions themselves have no time-based expiry. max_session_h left null.
- Hibernated-state storage, extra disk (EBS-gp3-like, unpublished), egress, IPv4 and concurrency limits are unpublished (null).
- Reserved capacity (15-min memory reservations, non-refundable, >=30 min lead) and Burst Sandboxes (~2x cheaper, restartable, alpha) exist in v1 docs but have no published rates, so they are not modeled as modes.
- Isolation: QEMU on KVM per the vendor's own guide. Region(s) unpublished.
- Benchmarks: DAX never succeeded; latest burst run failed (2026-09-25).
- Pro/Max included credit (10x fee) is shared between model tokens and machine time ("credit applies to both model calls (at API rates) and machine time"); the estimate assumes all of it goes to machine time, which is optimistic if the agent also spends tokens.
- SSH: none. Interactive access is a PTY "like SSH but over WebSocket" via the SDK (docs.opencomputer.dev/sandboxes/interactive-terminals). Desktop: none documented; only headless Chromium ("drive a browser").
- Fork (features check 2026-09-28): v1 docs say checkpoints let you "Fork new sandboxes from it ... like git branches for VMs" (disk-only or full disk+memory+CPU checkpoint). Docker-inside and /dev/kvm are not documented (null).
- Product pivot (2026-09): homepage now 'The platform for long-running agents' (serverless agents defined in TypeScript); docs.opencomputer.dev and its llms.txt cover only Serverless Agents, and the sandbox and browser-session doc pages there return 404. Sandbox docs survive only on docs-v1.opencomputer.dev (older runtime: 1/4/8/16 GB tiers). The sandboxes marketing page is the only current sandbox price source.
- Idle default: for bare SDK sandboxes the default timeout is 0 = never auto-hibernate, so a sandbox created with defaults bills until killed; idle hibernation needs an explicit timeout (e.g. 300 s). Agent sessions hibernate when idle by default (docs-v1.opencomputer.dev/sandboxes/timeout.md).

## How this provider charges


OpenComputer (diggerhq, Apache-2.0) is a managed runtime for long-running agents. It gives each session a QEMU/KVM
Linux VM, and it also has a sandbox SDK. Pricing has two meters: **model tokens**, passed through at API rates (free
with your own key or a Codex subscription), and **machine time** per second. Plans are prepaid credit at **10× the fee**.

Machine rates (pricing section of https://opencomputer.dev/, "billed to the second"):
1 vCPU / 2 GB **$0.00315/min = $0.189/h** (default) · 2 vCPU / 4 GB **$0.378/h** · 4 vCPU / 8 GB **$0.756/h**
(bursting is automatic). That is linear at $0.0945 per (1 vCPU + 2 GB)-hour. 4 vCPU / 8 GiB = **$0.756/h**.

## Regime table

| Regime | When it applies | How billed | Numbers | Source |
|---|---|---|---|---|
| PAYG | Default | Per-second machine time while running + tokens at API rates | rates above; $10 one-time free credit | https://opencomputer.dev/pricing |
| Pro | $20/mo | Prepaid: fee becomes 10× usage credit for both meters; PAYG after | $20 → $200 credit | same |
| Max | $200/mo | Same, "10+ default machines around the clock" | $200 → $2,000 credit | same |
| Enterprise | Own cloud/VPC, volume pricing, SSO/audit | Custom | null | same |
| Automatic burst | Session needs more RAM/CPU | Bills the burst tier while in it (v1 docs: bill = allocated memory tier × time) | ×2 / ×4 of default | pricing; https://docs-v1.opencomputer.dev/sandboxes/usage.md |
| Hibernated | Idle (sandboxes: 300 s rolling idle timeout) | VM state snapshotted; **no compute billed**; storage billing not stated | $0 compute | https://docs.opencomputer.dev/sandboxes/overview |
| Agent computer lifetime | Any agent session | One computer runs max 1 h, then it is replaced between commands; session has no time expiry | — | github.com/diggerhq/opencomputer/pull/806 |
| Disk beyond 20 GB | Closed beta, up to 256 GB | Per second "at a rate comparable to AWS EBS gp3" | rate unpublished | https://docs.opencomputer.dev/sandboxes/overview |
| Checkpoints (v2) | Filesystem only; `full` rejected | Storage billing not stated | null | https://docs.opencomputer.dev/sandboxes/checkpoints |
| Burst Sandboxes (alpha, v1) | Restart-tolerant workloads; up to 25 s notice; best-effort capacity | "roughly 2x cheaper than on-demand" | rate unpublished | https://docs-v1.opencomputer.dev/sandboxes/burst-sandboxes.md |
| Reserved capacity (v1) | Reserve memory GB per 15-min interval, ≥30 min ahead | `reservedGb × 900` GB-s per interval at the reserved rate, used or not; overage per second at on-demand; non-refundable, no carry-over | reserved rate unpublished | https://docs-v1.opencomputer.dev/reserved-capacity/usage-and-overage.md, /reserving.md |
| Self-host | Apache-2.0 repo + SELFHOSTING.md | Own infra | $0 licence | https://github.com/diggerhq/opencomputer |

### Dated change
Guides dated **2026-07-20** quoted "**$0.004/min flat = $0.24/hr**" for a **4 GB / 1 vCPU** default with RAM resizable 1–16 GB
(https://opencomputer.dev/guides/e2b-alternatives/). The current pricing is $0.00315/min for **2 GB / 1 vCPU**, bursting to
8 GB / 4 vCPU. A 4 GB machine now costs $0.378/h instead of $0.24/h, a 57% increase for that shape. The 1 vCPU price is 21% cheaper, but you get half the RAM.
PR #806 (2026-09-25) corrected an "eight-hour" computer lifetime that customers had been told to **one hour**.

## Gotchas

1. **Plans are 10× credit, not a fee**: Pro makes the first $200 of usage cost $20 (a 90% discount). Beyond that you pay list price.
2. **Credit is shared with tokens.** If you use OpenComputer's model keys, LLM spend eats the machine-time credit.
3. **Bursting is automatic.** A 4/8 workload is billed at $0.756/h only while it runs at that tier. Workloads that are mostly idle hibernate for free.
4. **Reserved capacity is use-it-or-lose-it per 15-min slot** and cannot be cancelled. Overage above the ceiling is billed per second at on-demand.
5. Storage, egress, concurrency limits and regions are all unpublished.
6. Doc sets are split (docs-v1 vs docs), so which v1 features (reserved capacity, burst sandboxes, memory checkpoints, forks) are still sold is unclear.

## Worked example

4 vCPU / 8 GiB (burst 4/8 tier), 50 concurrent × 8 h/day × 22 days = 8,800 machine-hours, 30% CPU,
50 GiB snapshots, 100 GiB egress. Hibernated between shifts. Tokens via own key ($0).

| Regime | Monthly total |
|---|---|
| PAYG, always at 4/8 tier | 8,800 × $0.756 = **$6,652.80** (+ unknown hibernated storage / egress) |
| Pro | $20 + (6,652.80 − 200) = **$6,472.80** |
| Max | $200 + (6,652.80 − 2,000) = **$4,852.80** |
| Max, autoscaled half the time at 1/2 default | $200 + (8,800 × 0.4725 − 2,000) = $2,358.00 |
| Enterprise / reserved capacity | unpublished |

CPU utilisation (30%) doesn't change the bill directly. Billing follows the allocated tier. Concurrency of 50 is not published as allowed.