# Prime Intellect Sandboxes

Official pricing: https://www.primeintellect.ai/sandboxes  
Category: agent-sandbox · Isolation: vm

## Pricing regimes (raw)

- **VM sandbox, launch pricing (valid through 2026-12-22)** (resource): $0.02/vCPU-h, $0.0125/GiB-h
- **GPU Pod marketplace, H100 80GB PCIe on-demand (cheapest live offer 2026-09-29: Lambda, 26 vCPU / 200 GB)** (resource): $0/vCPU-h, $0/GiB-h [stock]

## Features

Yes: idle auto-stop, ≥24 h sessions, custom image (Docker or snapshot), your own Docker/OCI image, full VM (own kernel), Docker inside, systemd, SSH, egress allowlist, open internet, SOC 2, GPU, EU data residency, spend limits, agent harness API, 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, nested virtualization, browser, desktop GUI, computer-use API, code interpreter, browser + desktop control, anti-bot stealth, CAPTCHA solving, residential IPs, preinstalled agents, public IPv4, HTTPS preview URLs, custom domains, raw TCP inbound, static egress IP, self-hosting, BYOC, open source, HIPAA, SSO, arm64, Windows, macOS, Python, Node.js, GPU desktop, credential injection, HTTP method/path egress rules, wake on request, live resize, live fork (no pause), memory fork, MCP server, automatic snapshots, snapshots on demand, 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/prime-intellect.json → feature_evidence

## Caveats

- Rates are PROMOTIONAL launch list prices 'valid through December 22, 2026' (docs). Post-promo pricing is unpublished.
- Billed on allocated vCPU/RAM/disk while running; 30% CPU utilisation saves nothing. Billing granularity and minimum are undocumented.
- No pause, snapshot, fork or persistent volume yet ('coming soon'). Keeping state means leaving the sandbox running (billed) or rebuilding from an image.
- Idle timeout (optional, 1-1,440 min) TERMINATES the sandbox; the default lifetime is 60 min (can be unlimited).
- disk_gib_month 0.146 = $0.0002/GiB-h x 730 h. The disk only exists while the sandbox runs, so this overstates part-time use. The default 5 GiB disk costs ~$0.001/h.
- The first VM launch of a new image converts it (PENDING for minutes, ~10 min per the verifiers docs); whether that time is billed is unknown.
- GPU sandboxes (0-8 GPUs, VM-only) are coming soon and need a grant; no GPU types or prices are published. Prime's separate GPU Pods marketplace is modelled as the stock-flagged 'gpu-pod-h100' mode at the homepage H100 listing ($2.43/hr); actual price depends on the provider picked in the dashboard.
- Hosted Evaluations on Prime Inference don't bill sandbox runtime separately (tokens only). Not modelled.
- No documented inbound port exposure; Prime Tunnel (unpriced) can publish an HTTPS URL. inbound_https is set false.
- Secrets are injected as env vars visible to the sandbox (not proxy-swapped), so credential_injection is false. Egress rules are domain/IP only.
- Preinstalled languages depend on the image you pick (default python:3.11-slim), so lang_* is null.
- Egress price, regions and compliance certifications are undocumented (null).
- Duplicate: cards/prime-intellect-sandboxes.json describes the same product with the same rates. Keep only one in the estimator.
- Replaces an earlier stub card that wrongly said 'container' isolation with a separate beta VM mode. Docs: 'VM sandboxes are the default runtime'.
- Launch pricing end date re-verified 2026-09-28 (docs sandboxes/overview: 'For our new launch, these rates are valid through December 22, 2026'); card price_expires = 2026-12-22 already set.
- Prepaid wallet is non-refundable: Terms of Service 'All sales are final and no refund will be issued.' (support may issue credits for provider failures per FAQ). GPU Pods are auto-deleted when the wallet runs out.
- GPU marketplace prices move in real time; the homepage GPU tiles are static marketing (e.g. H200 tiles showing '80 GB VRAM' at $0.47-$1.99/HR) and must not be used as rates. Use app.primeintellect.ai/api/trpc/availability.gpuSummary (public) for current cheapest per GPU.

## How this provider charges


Prime Sandboxes are hardware-isolated microVMs built for agentic RL. Each one boots a Docker/OCI image as the root filesystem
with its own guest kernel. The VMM isn't disclosed, and the blog positions the product against gVisor containers.
VM is now the default and only documented runtime. You drive sandboxes through the `prime` CLI, the `prime-sandboxes` Python SDK,
verifiers `SandboxEnv`, prime-rl, or Hosted Evaluations/Training. They went GA on 2026-09-23 (https://www.primeintellect.ai/blog/sandboxes),
after a beta that started 2025-08-27 with the Environments Hub.

There is one meter with no tiers, no subscription and no minimum spend. **All published rates are launch promo list prices valid
through 2026-12-22.** 4 vCPU / 8 GiB = 0.08 + 0.10 = **$0.18/h** (+$0.001/h for the default 5 GiB disk).

## Regime table

| Regime | When it applies | How billed | Numbers | Source |
|---|---|---|---|---|
| On-demand (launch promo) | Every running sandbox through 2026-12-22 | Allocated vCPU + RAM + disk while RUNNING; granularity/minimum undocumented | $0.02/vCPU-h, $0.0125/GiB-h RAM, $0.0002/GiB-h disk. Docs example 1 vCPU/2 GiB/5 GiB = $0.046/h; site calculator 1/1/5 = $0.0335/h | https://docs.primeintellect.ai/sandboxes/overview.md, https://www.primeintellect.ai/sandboxes |
| Post-promo | From 2026-12-23 | Unpublished | null | same |
| Default account limits | All accounts | Hard caps, raisable via support | 1,024 active sandboxes, 4,096 vCPU, 4,096 GiB RAM, 32,768 GiB storage, 102,400 creations/h, burst 192/10 s | overview.md |
| Custom capacity | Above the defaults ("tens of thousands" of concurrent sandboxes) | Contact | Unpublished | https://www.primeintellect.ai/blog/sandboxes |
| Per-sandbox size | Create request | Whole vCPUs; RAM and disk free-form | 1–16 vCPU, 128 MiB–64 GiB RAM, 2–128 GiB disk, 0–8 GPU (defaults 1 / 1 GiB / 5 GiB / 0) | overview.md |
| Lifetime / idle timeout | Always / opt-in | Sandbox **terminated** (not paused) | Default 60 min, negative = unlimited; idle 1–1,440 min, SSH doesn't count as activity | https://docs.primeintellect.ai/sandboxes/cli.md |
| GPU sandboxes | Not open | "Coming soon", VM-only, needs explicit grant | no types/prices published | overview.md |
| Pause / snapshot / fork / volumes | Not available | "Coming soon" | n/a | overview.md, blog |
| Hosted Evaluations (Prime Inference) | `prime eval` hosted, no `--api-base-url` | Tokens only; "sandbox runtime is not billed separately" | Prime Inference per-model token prices | https://docs.primeintellect.ai/tutorials-environments/hosted-evaluations |
| Hosted Evaluations (custom endpoint) | `--api-base-url` set | Billed as sandbox compute; the external provider bills tokens | rates above | same |
| Hosted Training (RL) | Managed RL runs | Per 1M tokens (input / output / train) | e.g. Qwen3.5-0.8B $0.02/$0.06/$0.06; Qwen3.6-35B-A3B $0.25/$0.75/$1.00; sandbox add-on undocumented | https://docs.primeintellect.ai/hosted-training/models-and-pricing |
| Egress | All traffic | Not published | null | (none found) |
| Payment | All | Prepaid Prime wallet (auto-top-up, usage limits, team wallets, promo codes) | n/a | https://docs.primeintellect.ai/faq |
| (Separate) GPU Pods marketplace | GPU VMs from aggregated providers, not sandboxes | Per-offer $/h, dynamic, deducted per minute; spot "up to 90%" off | Prices only visible logged in; docs example 2x H100 80GB PCIe (runpod) $5.40/h | https://docs.primeintellect.ai/faq, llms-full.txt |

## Gotchas

1. **The promo ends 2026-12-22.** Prime markets the launch rates as "a third of the cost of other large sandbox providers". Future rates are not published.
2. **There is no pause and no persistence.** Keeping state between shifts means leaving the sandbox running at full price. Otherwise you rebuild it from a Prime Image.
3. **The idle timeout deletes the sandbox.** An open SSH session doesn't keep it alive.
4. **Billing is on allocation.** Low CPU utilisation saves nothing.
5. **The first launch of any new image converts it to a VM image.** The sandbox stays PENDING for "a few minutes" (~10 min per the verifiers docs). Whether that time is billed is undocumented. Pre-pushing with `--source-image` avoids it.
6. **Concurrency for RL:** the default is 1,024 concurrent sandboxes, 4,096 vCPU and 4,096 GiB RAM. The vCPU cap binds first above 4 vCPU per sandbox (1,024 × 4 = 4,096). Thousands of rollouts need a support request.
7. **There is no inbound port API.** Prime Tunnel is the workaround and is unpriced. Uploads are capped at 16 MiB per file, and SSH has no port forwarding, scp or sftp.
8. **GPU sandboxes can't be bought yet.** The GPU Pods marketplace is a different product (GPU VMs with SSH) and does not support the sandbox API.

## Worked example

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

- Compute: 8,800 × $0.18 = **$1,584.00** (vCPU $704 + RAM $880). The 30% CPU figure is irrelevant because billing is on allocation.
- Disk: the default 5 GiB costs 8,800 × 5 × $0.0002 = **$8.80** (50 GiB disks would cost $88.00).
- Snapshots (50 GiB retained): **not possible** because the feature doesn't exist yet.
- Egress 100 GiB: the price isn't published, so it is treated as unknown ($0 in the total).

| Regime | Feasible? | Monthly total |
|---|---|---|
| Promo, sandboxes deleted after each 8 h shift | Yes (50 ≤ 1,024; 200 vCPU ≤ 4,096) | **$1,592.80** + unknown egress |
| Promo, kept running 24 h on 22 workdays to preserve state | Yes | 26,400 h × $0.181 = **$4,778.40** |
| Promo, kept running all month (730 h × 50) | Yes | 36,500 h × $0.181 = **$6,606.50** |
| After 2026-12-22 | ? | unknown |

## Sources

- https://www.primeintellect.ai/sandboxes
- https://www.primeintellect.ai/blog/sandboxes
- https://docs.primeintellect.ai/sandboxes/overview.md, cli.md, sdk.md, images.md, tunnel.md
- https://docs.primeintellect.ai/faq
- https://docs.primeintellect.ai/tutorials-environments/hosted-evaluations
- https://docs.primeintellect.ai/hosted-training/models-and-pricing
- https://raw.githubusercontent.com/PrimeIntellect-ai/prime/main/packages/prime-sandboxes/src/prime_sandboxes/models.py