# Tensorlake Sandboxes

Official pricing: https://www.tensorlake.ai/pricing  
Category: agent-sandbox · Isolation: firecracker

## Pricing regimes (raw)

- **Pro rate set (active CPU, allocated RAM/disk)** (resource): $0.042/vCPU-h, $0.009/GiB-h
- **BYOC / Self-hosted Compute (your AWS bill, reference rate, + 30% fee)** (resource): $0.094375/vCPU-h, $0/GiB-h [sales, alt]
- **Usage Credits prepaid rate set** (resource): $0.07/vCPU-h, $0.015/GiB-h
- **Free unmetered sandbox** (sizes): Free 1 vCPU/1 GiB $0/h

## 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), Docker inside, root, systemd, browser, desktop GUI, computer-use API, browser automation API, desktop control API, browser + desktop control, SSH, HTTPS preview URLs, egress allowlist, open internet, private networking, BYOC, SOC 2, HIPAA, SSO, GPU, Python, Node.js, wake on request, live fork (no pause), memory fork, webhooks, EU data residency, zero data retention, spend limits, automatic snapshots, snapshots on demand, fast boot (benchmarked), gVisor or VM (no shared kernel), inbound access rules, firewall inside (nftables), extra volumes, shared volumes

No: nested virtualization, code interpreter, anti-bot stealth, CAPTCHA solving, residential IPs, preinstalled agents, public IPv4, raw TCP inbound, static egress IP, self-hosting, open source, arm64, Windows, macOS, GPU desktop, HTTP method/path egress rules, live resize, MCP server, agent harness API, hosted agent API (their own agent), secret proxy, secret proxy for any API

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

## Caveats

- Engine encoding: rates are tied to plans (Pro rate set only on Pro, Usage Credits rate set only on the prepaid plan, Free unmetered only on Free) but the engine chooses modes independently of plans. The card therefore prices only the Pro path (mode pro + Pro plan = max($250, usage at Pro rates)); Free and Usage Credits are kept as trial_only plans with their rates in the notes. Workloads under ~$417/month at Credits rates (~$250 at Pro rates) are really cheaper on Usage Credits (min $5 pack) or Free; the estimate overstates them (floor $250). Root disk: Pro $0.0001/GB-h = storage.disk_gib_month $0.073; Usage Credits disk is 2x ($0.146).
- Active CPU falls back to allocated vCPU time for any sandbox generation where active metering is unavailable or untrustworthy. At 30% util that makes CPU 3.3x costlier. The engine assumes active metering works.
- RAM and root disk are billed on allocation while running, including the idle tail before timeout_secs (default 600 s, driven by proxied traffic).
- Snapshot storage $0.07/GB per fixed 30-day month, billed 'while suspended'. The price of snapshot artifacts and clones not tied to a suspended sandbox is not documented (assumed the same).
- Only named sandboxes can suspend. Ephemeral sandboxes terminate on idle (restartable for 48 h; billing in that window undocumented).
- Usage Credits has no postpaid overage: running sandboxes are stopped at zero balance. Pro bills overage at $0.01/CU.
- Plan session caps (2 h / 24 h / unlimited) are maximum idle timeouts, not wall-clock lifetimes.
- Warm-pool (warm_containers) billing is undocumented. Cloud Volumes, managed Git and GitHub Actions runner pricing are not on the sandbox rate card.
- BYOC mode = your cloud bill + Tensorlake fee: the estimate includes an AWS reference rate (i7i.metal-24xl $9.06/h, 96 vCPU, from the 2026-07-28 blog) as a stand-in for the AWS invoice you pay directly, plus the 30% entry fee. The fee falls to 10% (CPU) and to 2% (GPU) at scale. Excludes EBS and AWS egress. Your actual AWS pricing (reserved/savings plans, other instance types) changes the total.
- EU region (api.eu.tensorlake.ai): no price difference published.
- No GPU sandboxes on managed cloud (GPUs only in the separate Orchestration product and in BYOC at a 5-2% fee).
- Billed in whole CUs pooled per org (rounding at most $0.01). Pro CU rollover not stated.
- perf from benchmarks.json computesdk (cold start 706 ms, burst 1,127 ms, dax 97.2 s). No hpc CPU benchmark.
- Verified 2026-09-28 against https://www.tensorlake.ai/pricing ("UPDATED Q3 2026"): all rates, limits, packs, Pro $250 = 25,000 CU included, $0.01/CU overage, egress free, active-CPU fallback clause confirmed.
- Features check 2026-09-28: nested_virt left null. Docs say sandboxes are "MicroVMs backed by Firecracker and CloudHypervisor" (docs.tensorlake.ai/sandboxes/introduction.md); Cloud Hypervisor can expose KVM, so the Firecracker no-nested-virt rule cannot be applied, and /dev/kvm is not documented either way.
- Free tier session cap depends on verification: the API reference gives 'Free unverified 3600 (1h), Free verified 7200 (2h)' as timeout_secs maximums. The pricing page shows only 'Up to 2 hours'.
- No create/start rate limit is published. Creates beyond host, pool or namespace capacity are queued in Pending (pending_reason no_resources_available / pool_at_capacity / max_capacity_reached) and wait indefinitely unless max_pending_secs is set. The docs don't say whether Pending time is billed. 'When the fleet has to add a host first, that can take minutes.'
- The minimum root disk is 10 GiB (disk_mb 10240-102400), so every running sandbox bills at least 10 GB of disk ($0.001/h on Pro, $0.002/h on Credits) even if the workload asks for less.
- Cloud Volumes permanent snapshots are 'billed storage', but no rate is published on the pricing page or in the docs.
- Docs describe Pro as '$250 / month + usage' (billing table) and 'Pro is $250 per month plus metered usage' (pricing FAQ). The pricing page says the $250 'INCLUDES 25,000 CU ($250 OF USAGE)'. The card follows the pricing page (fee_is_credit true), and the docs FAQ defers to it ('For current commercial terms, see tensorlake.ai/pricing').
- GitHub Actions runners (separate product, not modelled) come as presets tensorlake-small 2 vCPU/4 GiB/10 GiB, medium 4/8/50, large 8/16/100 and xlarge 16/32/100, or custom sizes. Each runner has a 2 h maximum lifetime and draws on the project's sandbox quotas. No separate runner price is published (presumably the sandbox rate set).

## How this provider charges


Tensorlake (Firecracker/Cloud Hypervisor microVMs) meters **active CPU** (idle and I/O wait are free), plus **allocated** RAM and root disk while running, per second. Snapshot storage is billed while suspended. Everything is priced in compute units (1 CU = $0.01), pooled per org and billed in whole CUs.

The same workload lands on very different rates depending on:
- **the tier**. Usage Credits (prepaid) and Pro ($250/cycle, 40% lower rates) are separate rate sets (`usage_credits_gb_v1`, `pro_gb_v1`).
- **whether active-CPU metering works** for that sandbox generation. If it doesn't, CPU falls back to allocated.
- **named (suspendable) vs ephemeral** sandboxes.
- **BYOC**, where you pay your cloud plus a % fee.

## Regime table

| # | Regime | When it applies | How billed | Numbers | Source |
|---|---|---|---|---|---|
| 1 | Free | Default, no card | **Unmetered** (no charge) but tiny | $0. 1 sandbox at a time, 1 vCPU / 1 GB / 10 GB, idle timeout max 2 h, encrypted storage with 7-day retention | https://www.tensorlake.ai/pricing |
| 2 | Usage Credits (prepaid, "On-Demand") | Buy a $5/$10/$20 pack (500/1,000/2,000 CU). Up to 10 packs at once. Optional threshold auto-refill | Deducts from balance at the Usage-Credits rate set. **No postpaid overage**: at zero balance, new activity is blocked and running sandboxes are **stopped** | Active CPU $0.07/core-h, RAM $0.015/GB-h, disk $0.0002/GB-h, snapshot $0.07/GB-month. Limits: 100 concurrent, 4 vCPU / 16 GB / 100 GB, 24 h idle-timeout cap. 4/8 at 100% CPU = $0.40/h; at 30% = $0.204/h | https://www.tensorlake.ai/pricing, https://docs.tensorlake.ai/platform/billing |
| 3 | Pro | $250 per monthly billing cycle | Fee = 25,000 CU of usage **at Pro's lower rates**. Overage $0.01/CU on the same invoice. So you pay max($250, usage) | Active CPU $0.042, RAM $0.009, disk $0.0001/GB-h, snapshot $0.07/GB-month. Limits: 1,000 concurrent, 16 vCPU / 64 GB / 100 GB, unlimited duration, HIPAA BAA, 24×7 Slack. 4/8 at 100% = $0.24/h; at 30% = $0.1224/h | https://www.tensorlake.ai/pricing |
| 4 | Enterprise | Quote | Custom rates, unlimited concurrency, custom vCPU/RAM/disk ceilings, SSO/SAML/RBAC, in-VPC/on-prem, 1 h P1 SLA | null | https://www.tensorlake.ai/pricing |
| 5 | BYOC / Self-hosted Compute | Project with Self-hosted Compute enabled. Sandboxes run on your AWS (i7i.metal executors) | You pay AWS directly. Tensorlake charges a **% of compute cost**: CPU sandboxes 30%, down to 10% at scale; GPU sandboxes 5%, down to 2%. Blog example: i7i.metal-24xl (96 vCPU / 768 GiB) $9.06/h at a 20% rate = $10.87/h, "~1.1¢ per sandbox-hour" at 1,000 sandboxes/machine | 30%→10% (CPU), 5%→2% (GPU) | https://www.tensorlake.ai/blog/introducing-tensorlake-byoc (dated 2026-07-28), https://docs.tensorlake.ai/platform/self-hosted-compute/aws |
| 6 | Active-CPU metering | Normal case | Only core-time actually executing is billed. Idle and I/O wait are free. RAM and disk are billed on allocation regardless | cost ∝ utilisation (CPU only) | /pricing FAQ, /docs/platform/billing |
| 7 | Allocated-CPU **fallback** | "If an active CPU sample is unavailable or untrustworthy" for a sandbox generation | CPU billed on **allocated** vCPU time for that generation | Pro 4/8 = $0.24/h. Usage Credits = $0.40/h | https://www.tensorlake.ai/pricing (FAQ "How is pricing metered?") |
| 8 | Running-idle tail | Until `timeout_secs` idle threshold (default **600 s**). Driven by proxied traffic (SDK, SSH, PTY, exposed ports) | RAM + disk at full rate. CPU ≈ $0 because it is idle | Pro: 8 GB × 0.009 = $0.072/h. Credits: $0.12/h | https://docs.tensorlake.ai/sandboxes/lifecycle |
| 9 | Suspended (**named sandboxes only**) | Manual suspend or idle timeout on a named sandbox | "Consumes no compute". Snapshot storage **per GB per fixed 30-day month** | $0.07/GB-month (both tiers) = $0.0000972/GB-h | https://docs.tensorlake.ai/sandboxes/lifecycle, /pricing |
| 10 | Ephemeral (unnamed) sandboxes | Default. Cannot suspend | Terminated on idle timeout. No storage after that | $0 | https://docs.tensorlake.ai/sandboxes/lifecycle |
| 11 | Terminated, restart window | 48 h after termination (restores from latest snapshot if any) | Billing during the window **not documented** | null | https://docs.tensorlake.ai/sandboxes/lifecycle |
| 12 | Snapshot artifacts (filesystem / memory) and `copy -n N` clones | Snapshots used to create new sandboxes or clones | Docs give **no price**. The pricing page only prices snapshot storage "while suspended". Presumably the same $0.07/GB-month | null (assume 0.07) | https://docs.tensorlake.ai/sandboxes/snapshots |
| 13 | Warm pools | `warm_containers` buffer per pool | Pre-booted idle containers. **Billing undocumented**. Likely RAM + disk while warm, since CPU is idle | null | https://docs.tensorlake.ai/sandboxes/pools |
| 14 | Egress | Any | **Free** ("Network egress is free") | $0 | https://www.tensorlake.ai/pricing |
| 15 | EU region | `api.eu.tensorlake.ai` | Same API key. **No price difference published** | multiplier 1 | https://docs.tensorlake.ai/platform/eu-data-residency |
| 16 | Rounding | Always | Measured per second, usage pooled per org, **whole CUs** ($0.01) billed | ≤ $0.01 per org per bill | /pricing, /docs/platform/billing |
| 17 | Plan switching | Credits ↔ Pro | Unused credits are retained (not refunded) and return if you go back. Going back to Credits **requires buying a new pack**. Pro CU rollover is not stated (assume none) | | /pricing FAQ |
| 18 | Cloud Volumes, managed Git, GitHub Actions runners, Document AI | Adjacent products | Not on the sandbox rate card. Document AI draws from the same credit pool | null | https://docs.tensorlake.ai/filesystems/introduction |

## Gotchas

1. **CPU is cheap, RAM is the bill.** On Pro, 8 GB RAM ($0.072/h) costs more than 4 fully busy cores × 30% ($0.050/h). RAM must be 1-8 GB per vCPU, so you cannot drop RAM below 1 GB/vCPU.
2. **Two rate cards, 40% apart.** The Pro rate set only applies on Pro. The $250 fee is fully spendable, so Pro wins once monthly usage at *Credits* rates exceeds ≈ $417 (250 / 0.6). The calculator's own break-even message is "≈ 5,661 sessions of 25 min, 2 vCPU/4 GB".
3. **Usage Credits is a hard cap, not a budget alert.** Hit zero and running sandboxes are **stopped**. Auto-refill is opt-in.
4. **Session limits are idle limits.** The "2 h / 24 h / unlimited" plan caps are the maximum `timeout_secs` idle threshold, not wall-clock lifetimes (the pricing page calls them "per sandbox session").
5. **Only named sandboxes suspend.** Ephemeral ones are terminated on idle, so their state is lost (48 h restart window aside).
6. **The fallback clause.** If active metering is "unavailable or untrustworthy", CPU reverts to allocated. That makes CPU 3.3× costlier at 30% util, and you can't predict when it happens.
7. **Snapshot storage at $0.07/GB-month** is on the high side, and it uses a fixed 30-day month (≈ 1.4% more than per-730 h). It is charged at the same rate on both tiers.
8. **Usage Credits caps a sandbox at 4 vCPU / 16 GB.** Larger needs Pro.
9. **BYOC is a %-of-cloud-bill fee**, so it scales with your AWS instance prices. At the entry 30% CPU rate the fee plus AWS may exceed managed Pro unless you pack densely.

## 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.

Assumptions:
- Disk is 10 GB per sandbox = 88,000 GB-h.
- Active core-h = 8,800 × 4 × 0.3 = 10,560. RAM GB-h = 70,400.
- The 50 GiB snapshot is billed at $0.07/GB-month = $3.50.
- Egress is free.
- 50 concurrent fits Usage Credits (100) and Pro (1,000). 8 h sessions need idle timeouts ≤ 24 h, which is OK on both.

| Regime | Active CPU | RAM | Disk | Snapshots | Egress | Plan fee | **Monthly total** |
|---|---|---|---|---|---|---|---|
| Usage Credits, active metering | 10,560 × 0.07 = $739.20 | 70,400 × 0.015 = $1,056.00 | $17.60 | $3.50 | $0 | $0 (prepaid ≈ 91 × $20 packs) | **$1,816.30** |
| **Pro, active metering** | 10,560 × 0.042 = $443.52 | 70,400 × 0.009 = $633.60 | $8.80 | $3.50 | $0 | $250 credit consumed | **$1,089.42** (= $250 + $839.42 overage) |
| Pro, allocated-CPU fallback | 35,200 × 0.042 = $1,478.40 | $633.60 | $8.80 | $3.50 | $0 | credit | **$2,124.30** |
| Usage Credits, allocated fallback | 35,200 × 0.07 = $2,464.00 | $1,056.00 | $17.60 | $3.50 | $0 | | **$3,541.10** |
| Idle-tail add-on (default 10 min idle per sandbox-day = 183.3 h) | ≈ $0 | | | | | | Pro +$13.38; Credits +$22.37 |
| Free | n/a (1 concurrent, 1 vCPU / 1 GB) | | | | | | n/a |
| Enterprise | custom | | | | | | null |
| BYOC, illustrative (blog's i7i.metal-24xl $9.06/h, 176 h) | 1 host, packed (60 active vCPU, 400 GiB fits 96/768): $1,594.56 AWS | | | | AWS egress extra | fee 30% / 10% | **$2,072.93 / $1,754.02** + EBS/egress |
| BYOC, no overcommit (200 vCPU → 3 hosts) | $4,783.68 AWS | | | | | fee 30% / 10% | **$6,218.78 / $5,262.05** |

- Pro is cheapest here: $1,089.42, which works out to about $0.124 per sandbox-hour.
- The 30% utilisation cuts the Pro CPU line by 70%.
- RAM is 58% of the Pro bill.
- BYOC rows are illustrative only. They depend on your instance choice, overcommit and AWS pricing, and exclude EBS and AWS egress.

Sources: https://www.tensorlake.ai/pricing · https://docs.tensorlake.ai/platform/billing · https://docs.tensorlake.ai/sandboxes/lifecycle · https://docs.tensorlake.ai/sandboxes/snapshots · https://docs.tensorlake.ai/sandboxes/pools · https://docs.tensorlake.ai/platform/eu-data-residency · https://docs.tensorlake.ai/platform/self-hosted-compute/aws · https://www.tensorlake.ai/blog/introducing-tensorlake-byoc