# OpenReward Sandboxes

Official pricing: https://openreward.ai/pricing  
Category: agent-sandbox · Isolation: container

## Pricing regimes (raw)

- **Pay-as-you-go, per second** (resource): $0.0504/vCPU-h, $0.0162/GiB-h

## Features

Yes: volumes, idle auto-stop, ≥24 h sessions, custom image (Docker or snapshot), your own Docker/OCI image, egress allowlist, open internet, self-hosting, GPU, Python, Node.js, credential injection, secret proxy, secret proxy for any API, extra volumes, shared volumes

No: snapshots, memory snapshots, fork/clone, pause/resume, persistent disk, start from your own snapshot, full VM (own kernel), Docker inside, nested virtualization, browser, desktop GUI, computer-use API, code interpreter, browser + desktop control, anti-bot stealth, CAPTCHA solving, residential IPs, preinstalled agents, SSH, public IPv4, HTTPS preview URLs, static egress IP, open source, arm64, Windows, macOS, GPU desktop, HTTP method/path egress rules, wake on request, live resize, live fork (no pause), memory fork, 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)

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

## Caveats

- Sandboxes are tied to an OpenReward environment workspace ('not a general sandbox compute solution'). Usable for RL/eval environments, not as a general agent VM.
- Only preset CPU:RAM sizes up to 4 vCPU / 16 GiB; no larger machines.
- The nvidia-l4 preset's vCPU/RAM and whether CPU/RAM are billed on top of $1.51/GPU-h are undocumented; the engine adds the GPU rate on top of resource rates.
- Concurrency limits, max lifetime, minimum billed duration, disk and egress pricing are undocumented (null).
- No pause/snapshot; the network option is all-or-nothing block_network (no allowlist).
- CPU/RAM rates are numerically identical to E2B's list rates; the backend is not disclosed.
- Academic credits are available on request (not modelled).
- SSH: not offered in the docs (sandbox concepts/openreward pages, llms.txt index list no SSH); access is through the environment session SDK. Disk size and disk price are not published (pricing page lists CPU, memory and L4 GPU only), so extra disk stays unpriced (2026-09-28).
- Hosted OpenReward environments bill an environment server in addition to the sandbox. Published examples use 1 vCPU/4 GiB = $0.0000320/s ($0.1152/h) at the public CPU+RAM rates. For example, JobTasks with a 2 vCPU/2 GiB sandbox totals $0.2484 per session-hour.
- Machine sizes: docs (concepts/sandboxes) list 0.5:0.5 to 4:16 plus nvidia-l4. The official Python SDK added 8:8, 16:16, 8:16, 8:32 and 16:64 in v0.1.159 (2026-09-24), and backend availability of these sizes isn't documented. A separate docs page (sandboxes/openreward) claims sizes 'From 0.1:0.1', but no 0.1 size exists in the SDK or the size list.
- Custom sandbox images must be public on Docker Hub and built for linux/amd64 (no private registries, no arm64).
- Hosted-environment server sizes vary by environment: 1 vCPU/4 GB is the default (CLI default --cpu-memory 1:4), and 2:4 and 4:8 have also been seen. Each is billed per second of each session at the same CPU/RAM rates. Examples: DataScienceComps environment $0.0000920/s plus sandbox $0.0000640/s = $0.5616 per session-hour; KernelBench environment $0.0000460/s plus a task-dependent sandbox.

## How this provider charges


OpenReward (https://openreward.ai) is an RL environment hub built by GR Inc (General Reasoning), with 380+ hosted ORS environments.
Its native sandboxes are isolated containers provisioned through an environment workspace. The docs say they are "not a general
sandbox compute solution: it is meant to be tied to environment use". E2B, Daytona and Modal are documented as alternative backends.

## Regime table

| Regime | When it applies | How billed | Numbers | Source |
|---|---|---|---|---|
| Pay-as-you-go CPU sandbox | Any sandbox from `start()` to `stop()` | Per second on the chosen `machine_size` preset (allocated) | $0.0504/vCPU-h, $0.0162/GiB-h; presets 0.5:0.5 … 4:16 | https://openreward.ai/pricing, https://docs.openreward.ai/concepts/sandboxes.md |
| GPU sandbox | `machine_size="nvidia-l4"` | Per second | NVIDIA L4 $1.51/GPU-h (whether CPU/RAM are billed on top is undocumented) | pricing page, https://docs.openreward.ai/environments/gpu-environments.md |
| Hosted environment used by others | Someone runs your hosted environment | Sandbox time billed to the caller's account, not the host's | — | concepts/sandboxes.md |
| BackSearch / BackFetch | Point-in-time web tools | Per request | $10 / 1,000 searches; $2 / 1,000 fetches | pricing page |
| Researcher credits | Academic program | Credits on request | unpublished | pricing page |
| Enterprise | On-prem, dedicated support, SLAs | Contact sales | unpublished | pricing page |
| Disk / egress / snapshots | — | Not published | null | — |

## Gotchas

1. **The sandboxes are tied to environment workspaces.** You need an OpenReward environment namespace to get sandbox compute at all, so this isn't a general agent-VM product.
2. **The largest CPU size is 4 vCPU / 16 GiB,** and only fixed ratios are offered (1, 2 or 4 GiB per vCPU).
3. **There is no pause or snapshot.** Network control is all-or-nothing (`block_network`).
4. **Concurrency limits, max lifetime and minimum billed time are undocumented.**
5. **The CPU/RAM rates match E2B's list rates exactly.** The backend isn't disclosed.

## Worked example

4 vCPU / 8 GiB ("4:8" preset), 50 concurrent × 8 h/day × 22 days = 8,800 sandbox-hours, 30% CPU, 50 GiB snapshots, 100 GiB egress.

- Compute: 8,800 × (4 × 0.0504 + 8 × 0.0162) = 8,800 × $0.3312 = **$2,914.56** (vCPU $1,774.08 + RAM $1,140.48). Billing is on allocation, so the 30% CPU figure doesn't change it.
- Snapshots: not available. Egress: price not published (unknown).
- **Total ≈ $2,914.56/month** + unknown egress/disk. Whether 50 concurrent sandboxes are allowed is not documented.

## Sources

- https://openreward.ai/pricing (raw: research/raw/openreward_ai_pricing_2026-09-28.html)
- https://docs.openreward.ai/concepts/sandboxes.md
- https://docs.openreward.ai/sandboxes/openreward.md
- https://docs.openreward.ai/environments/gpu-environments.md
- https://docs.openreward.ai/llms-full.txt