# Novita AI Agent Sandbox

Official pricing: https://docs.novita.ai/guides/sandbox-pricing  
Category: agent-sandbox · Isolation: firecracker

## Pricing regimes (raw)

- **Pay-as-you-go (prepaid balance)** (resource): $0.03528/vCPU-h, $0.01152/GiB-h
- **GPU Instance on-demand (H100 SXM 80GB, L40S, RTX 4090, RTX 5090, RTX 6000 Ada)** (resource): $0/vCPU-h, $0/GiB-h [stock]
- **GPU Instance spot (H100 SXM 80GB, RTX 4090, RTX 6000 Ada)** (resource): $0/vCPU-h, $0/GiB-h [spot, stock]

## 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, your own Docker/OCI image, full VM (own kernel), root, systemd, browser, desktop GUI, computer-use API, code interpreter, browser + desktop control, preinstalled agents, SSH, HTTPS preview URLs, egress allowlist, open internet, SOC 2, Python, Node.js, credential injection, wake on request, live resize, memory fork, audit logs, automatic snapshots, snapshots on demand, 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, extra volumes, shared volumes

No: Docker inside, nested virtualization, anti-bot stealth, CAPTCHA solving, residential IPs, public IPv4, raw TCP inbound, static egress IP, private networking, self-hosting, BYOC, open source, HIPAA, GPU, arm64, Windows, macOS, GPU desktop, HTTP method/path egress rules, live fork (no pause), EU data residency, spend limits, MCP server, agent harness API

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

## Caveats

- Isolation technology not documented (E2B-derived envd, memory snapshots, virtio-mem suggest a microVM; unverified).
- Persistent storage (paused sandboxes + templates + snapshots) has a 60 GB free allowance per ACCOUNT (extra key snapshot_free_gib_account), then $0.00009/GB-h = $0.0657/GB-month; billed daily; landing (2026-09-29): 'Measured hourly and charged daily. First 60 GB included'. An engine ignoring the allowance overstates small snapshot bills.
- Paused sandbox state includes memory; its GB size is not documented.
- Max session: 1 h default cap without long_running; Paid 3 h (docs quota table) vs 24 h (landing); long_running examples 168 h / 720 h.
- Default timeout 5 min and default onTimeout action is kill.
- The $100 promo credit may not unlock Paid-tier limits; Paid requires a top-up or credit line. Credit validity period unspecified.
- No minimum billed duration or per-start fee documented (Novita blog: 'no additional per-session startup fee').
- Egress, volume (beta) and Enterprise pricing not published.
- Overdue prepaid balance terminates sandboxes with data loss; no spend caps.
- Ephemeral disk fixed at 20 GB (Enterprise adjustable); no per-GB disk upgrade price.
- Benchmarks: only hpc CPU (13.52 Node.js tooling runs/s, n=10); no computesdk cold-start/burst/dax data.
- Idle auto-stop exists but is off by default (must be configured), so the estimator does not apply idle suspension; set a pinned offer if you enable it.
- GPU Instances (novita.ai/gpus) are a separate GPU-cloud product from the CPU Agent Sandbox, modelled as alt-flagged gpu_only modes (patch v4-gpu, 2026-09-28). PPIO (ppio-sandbox card, China, CNY 0.108/vCPU-h + 0.054/GiB-h) is a separately billed offer; engineering overlap with Novita is visible (novita-gym PPIO templates) but common ownership / shared credits are unverified (patch v10-regional).
- GPU Instance regimes un-flagged alt (fix 2026-09-28): an ordinary self-serve GPU VM product; prices re-rendered from novita.ai/gpus 2026-09-28 (H100 SXM 3.39 / spot 1.70 /hr/GPU). Egress is still not published on docs.novita.ai/guides/sandbox-pricing or novita.ai/sandbox (egress_gib stays null).
- GPU Instance billing: 'accurate to the second and settled hourly'; billing starts when the instance 'enters the pulling state' (image pull time is billed, boot_overhead_s unknown). Stopped GPU instances keep data up to 7 days and pay $0.005/GB/day storage; container disk has 60 GB free then $0.005/GB/day; volume disk $0.005/GB/day; network volume $0.002/GB/day. These GPU storage rates are not modelled in the card storage block (which describes the Agent Sandbox).
- Agent Sandbox account quota also has aggregate limits: concurrentVcpu and concurrentRamMb (docs example for a 100-instance account: 200 vCPU / 409,600 MiB). Per-tier values are not published; if the Paid tier really is 200 vCPU, 100 concurrent sandboxes only fit at <=2 vCPU each (50 x 4 vCPU, 25 x 8 vCPU).
- GPU Instance prices vary by cluster in the first-party market API (not on rendered pricing pages): RTX 4090 $0.67/h (spot 0.335) in the Dallas clusters vs $0.33 elsewhere; RTX 5090 'RTX 5090 32GB' SKU $0.69-0.73/h (spot 0.35-0.365) vs the $0.56 'NVIDIA GeForce RTX 5090' SKU; L40S spot $0.275 appears in CN-HK-01 only. All of these were out of stock on 2026-09-29.
- GPU Instance overview advertises 'free network transfer'; Agent Sandbox egress is still unpublished (network.egress_gib null).

## How this provider charges


Pure prepaid pay-as-you-go: no subscription, no plan fee, no per-start fee. One compute rate card
($0.0000098/vCPU-s, $0.0000032/GiB-s, allocated, per-second). The regimes differ by **account tier
(quota unlock)**, **lifecycle state (running / paused / killed)**, **timeout mode (default vs long-running)**
and **storage (ephemeral vs persistent w/ 60 GB account allowance)**.

## Regime table

| Regime | When it applies | How billed | Numbers | Source |
|---|---|---|---|---|
| Running (on-demand) | Sandbox in Running state, any tier | Allocated vCPU + allocated RAM, per second (docs: "accurate to the second"); utilisation irrelevant | vCPU $0.0000098/s = **$0.03528/vCPU-h**; RAM $0.0000032/GiB-s = **$0.01152/GiB-h**; 4 vCPU/8 GiB = **$0.23328/h**; 8/8 = $0.3744/h (vendor example matches) | https://docs.novita.ai/guides/sandbox-pricing , https://novita.ai/sandbox |
| Size grid | Every sandbox | vCPU integer 1–8; RAM 512 MiB steps, 512–8192 MiB, 0.5–4 GiB per vCPU | 4 vCPU allows 2–8 GiB; 8 vCPU needs ≥4 GiB | https://docs.novita.ai/guides/sandbox-quota-limit |
| Free tier (no balance) | New account, no top-up / credit line | Same rates, drawn from $100 promo credit | 5 concurrent, **1 h** max session, max 2 vCPU / 4 GiB, 20 GB disk, no priority scheduling | quota-limit page, novita.ai/sandbox |
| Paid tier (PAYG) | Unlocks when account balance > $0 (landing) / "adding account balance or enabling a credit line" (docs) | Same rates, prepaid balance | 100 concurrent, max 8 vCPU / 8 GiB, 20 GB disk; session **3 h** (docs quota table) vs **24 h** (landing page); priority scheduling | quota-limit page, novita.ai/sandbox |
| Enterprise | Contract | Not published | 500 concurrent default (landing: "unlimited"), 4 h session default, all limits adjustable, region selection / dedicated infra, 99.95% SLA | quota-limit page, novita.ai/sandbox |
| Long-running mode | `metadata.long_running="true"` / CLI `--long-running` at create | Same running rates; only lifts timeout cap | Default timeout 5 min; cap 1 h without long-running; docs examples 24 h, 168 h, 720 h. Inherited through resume/clone | https://docs.novita.ai/guides/sandbox-long-running |
| Paused (manual / onTimeout=pause / idle auto-pause) | Sandbox paused; memory+FS preserved | CPU & RAM = $0; state counts as Persistent Storage | Persistent storage: **60 GB free per account**, then **$0.00009/GB-h** (= $0.0657/GB-month at 730 h; $0.00216/GB-day); "billed daily" (docs), "first 60 GB included daily" (landing); metering interval not documented | sandbox-pricing, novita.ai/sandbox, sandbox-overview |
| Auto-pause / auto-resume ("on-demand mode", NovitaClaw) | lifecycle onTimeout=pause + autoResume, or idle_timeout (60–86400 s in NovitaClaw) | Compute only while running; resumes (~1 s) on SDK call or inbound HTTP | Zero compute while paused; storage as above | llms-full (NovitaClaw), sandbox-idle-timeout |
| Templates & snapshots | Any built template, snapshot (v2 region only), legacy commit/clone snapshot-templates | Persistent Storage, same pool as paused sandboxes | Shares the **same 60 GB account allowance**, then $0.00009/GB-h | sandbox-pricing ("Templates and Snapshots are saved as persistent storage"), sandbox-overview |
| Ephemeral disk | Running sandbox | Free | 20 GB per sandbox, not expandable except Enterprise | sandbox-pricing |
| Volumes (beta, v2 only) | Standalone mounted volumes | **Not published** | null | https://docs.novita.ai/guides/sandbox-volume |
| Killed | After kill / timeout-kill (default) | Nothing | $0; state lost | sandbox-overview |
| Egress / ingress | Internet traffic | **Not published** (no line item on pricing page) | null (assumed $0, unverified) | sandbox-pricing |
| Regions | v2 us-phx-1 (default) vs v1 us-virginia-1 | No price difference documented | Snapshots, secrets, volumes only in v2 | sandbox-overview |
| Promo credit | New account after onboarding survey | One-time $100, applied first to CPU/RAM/storage | Validity period unspecified; "subject to availability" | sandbox-pricing |
| Overdue | Balance (incl. vouchers) < bill | Sandboxes **terminated, data lost** | — | sandbox-pricing |

## Gotchas

1. **Paying doesn't upgrade limits by itself.** The $100 promo credit is spent under Free-tier quota (2 vCPU / 4 GiB,
   5 concurrent, 1 h). Paid-tier limits (8/8, 100 concurrent) need a real balance top-up or credit line. Whether voucher
   balance counts as "balance > $0" is unclear.
2. **Session cap is contradictory**: 1 h default timeout cap (without long_running), 3 h (docs quota table, Paid),
   24 h (landing page), 720 h (long-running examples). With no minimum and no start fee, splitting sessions costs nothing extra.
3. **Default timeout is 5 minutes, and the default onTimeout action is kill.** Forget to set timeout → work lost; forget
   to pause → you pay until the timeout fires (up to 1 h, or whatever long-running timeout you set).
4. **The 60 GB free is per account, not per sandbox**, and templates + snapshots + paused sandboxes all draw from it.
   A few fat templates can eat the allowance so every paused sandbox is billed from byte 0. Paused state includes memory
   (pause takes ~4 s per GB RAM per legacy docs), so an 8 GiB sandbox pauses into ≥8 GB plus written disk.
5. **CPU utilisation is irrelevant**: CPU and RAM billed on allocation. A 30%-busy sandbox costs the same as 100%.
6. **Memory hotplug** (virtio-mem) changes allocated RAM of a running sandbox; billing presumably follows allocation
   (RAM "billed based on allocated memory capacity"), unverified.
7. **Novita's own blog (2026-06-30) misquotes RAM as $0.0000016/GiB/s** — that is the 512 MiB price. Official docs:
   $0.0000032/GiB/s.
8. **Hard prepaid cutoff**: going overdue kills sandboxes and deletes data; only email low-balance alerts, no spend caps.
9. Egress and volume pricing are not published; enterprise pricing is contract-only.
10. 8 GiB RAM max per sandbox on standard tiers; 20 GB disk hard limit (Enterprise adjustable).

## Worked example

Workload: 4 vCPU / 8 GiB, 50 concurrent × 8 h/day × 22 days = **8,800 sandbox-hours**, 30% CPU util,
50 GiB snapshots retained, 100 GiB egress. 4/8 is a legal size (2 GiB/vCPU). 50 concurrent ≤ 100 (Paid).

| Regime | Compute | Storage | Egress | Monthly total |
|---|---|---|---|---|
| Paid PAYG, kill after each session | 8,800 × $0.23328 = **$2,052.86** (30% util gives no discount) | 50 GiB ≤ 60 GB free → **$0** (if templates don't consume the allowance) | not published → $0 assumed | **$2,052.86** |
| Same, first month with $100 promo credit | — | — | — | **$1,952.86** (credit is one-time) |
| Same, but the 60 GB allowance already used by templates | $2,052.86 | 50 × $0.0657 = $3.29 | $0? | **$2,056.15** |
| Paid PAYG, pause instead of kill between sessions (illustrative assumption: 10 GB state per paused sandbox; not a Novita number) | $2,052.86 | (50×10 + 50 − 60) = 490 GB × 554 paused h × $0.00009 = $24.43 | $0? | **≈$2,077.29** |
| Free tier | not feasible: max 2 vCPU / 4 GiB, 5 concurrent, 1 h sessions | — | — | n/a |
| Enterprise | rates not published | — | — | null |

Session-length note: 8 h sessions need long-running mode (without it the timeout is capped at 1 h, so the
alternative is 8 × 1 h sessions — the 3 h Paid "maximum session duration" in the quota table does not help
without long-running; verified 2026-09-28 at https://docs.novita.ai/guides/sandbox-long-running);
cost is identical because there is no minimum or start fee.

Sources: https://docs.novita.ai/guides/sandbox-pricing , https://docs.novita.ai/guides/sandbox-quota-limit ,
https://docs.novita.ai/guides/sandbox-long-running , https://docs.novita.ai/guides/sandbox-overview ,
https://docs.novita.ai/guides/sandbox-volume , https://novita.ai/sandbox ,
https://blogs.novita.ai/ai-agent-sandbox-pricing-models/ (2026-06-30), research/raw/docs_novita_ai_llms_full.txt