<!-- markdownlint-disable MD013 -->

# sandboxprices.com default pricing comparison

> Agent-readable, source-linked monthly sandbox pricing generated from the same canonical data and calculation code as the interactive calculator.

Pricing audited as of: 2026-09-04

Canonical document: <https://sandboxprices.com/pricing.md>

Interactive calculator: <https://sandboxprices.com/>

API: <https://api.sandboxprices.com/api/tco>

This is not a generic hourly-price list. Every numeric value below is the calculator result for the exact canonical workload. `gross_usd_month` is the cross-provider comparison basis: provider resource meters before optional account plan fees, recurring credits, and account-wide allowances. `modeled_invoice_usd_month` applies the cheapest eligible public plan, required fee, recurring credit, allowance, or minimum that the calculator can reproduce. A product whose only public price is an inseparable fixed plan or bundle retains that inherent charge in both columns. Use the API or interactive calculator for any other workload.

A missing total is intentional. It means the public evidence cannot support a reproducible result or the default workload is unsupported. Missing components are never replaced with guessed market rates.

## Canonical workload

This request is anchored to AWS Lambda MicroVM's smallest published shape. Other products must accept it or round up to their smallest valid single-sandbox allocation.

```json
{
  "vcpus": 0.25,
  "ramGiB": 0.5,
  "diskGiB": 5,
  "meteredRamGiB": 0.5,
  "hotDiskGiB": 5,
  "coldDiskGiB": 5,
  "activeHoursPerMonth": 200,
  "suspendedHoursPerMonth": 50,
  "executionsPerMonth": 60,
  "resumesPerMonth": 50,
  "newSandboxesPerMonth": 60,
  "snapshotGiB": 2,
  "imageGiB": 2,
  "cpuUtil": 0.5,
  "burstX": 1,
  "cpuBurstX": 1,
  "ramBurstX": 1
}
```

Field interpretation:

- `activeHoursPerMonth` is aggregate running sandbox-hours, not a 730-hour reservation.
- `meteredRamGiB`, `hotDiskGiB`, and `coldDiskGiB` apply only when a provider meters observed use or retained bytes.
- `cpuUtil` applies only to providers that meter actual CPU consumption.
- `cpuBurstX` and `ramBurstX` independently describe Lambda use above its configured baseline.
- `executionsPerMonth`, `newSandboxesPerMonth`, `resumesPerMonth`, and `suspendedHoursPerMonth` drive per-run minimums and lifecycle charges.
- AWS EC2 `t4g.nano` is a permanent raw-VM comparison baseline. It receives the same workload, bills 730 always-on hours plus retained gp3 storage, and remains visible in the human calculator even when other self-hosted references are hidden.

## AWS Lambda MicroVM baseline result

```json
{
  "productId": "aws-lambda-microvms",
  "grossUsdMonth": 7.197115111,
  "modeledInvoiceUsdMonth": 7.197115111,
  "invoiceContext": "same as gross",
  "billedAllocation": {
    "vcpus": 0.25,
    "ramGiB": 0.5,
    "diskGiB": 5,
    "shape": "micro-0.25"
  },
  "componentsUsd": {
    "cpu": 4.984992,
    "memory": 1.320012,
    "activeDisk": 0,
    "retainedDisk": 0,
    "imageStorage": 0.16,
    "suspendedStateStorage": 0.011111111,
    "snapshotCycles": 0.535,
    "launchImageReads": 0.186,
    "creationAndResumeFees": 0
  }
}
```

- compute: CPU 0.09969984/vCPU·hr × 0.25 × 200h + RAM 0.01320012/GiB·hr × 0.5 × 200h
- base image storage: 0.08/GB·mo × 2 GB
- suspended state storage: 0.08/GB·mo × 2 GB × 50/720
- suspension duration: 1.000h implied average ≤ 8h limit
- snapshots: (0.0038/GiB write + 0.00155/GiB read) × 2 GiB × 50 resumes
- launch image reads: 0.00155/GB read × 2 GB × 60 new boxes
- lifecycle eligibility: 4.167h implied average running-plus-suspended lifetime ≤ 8h limit; resumes do not reset it
- bills the baseline while running, then independently meters additional CPU and memory up to 4×; inputs are 1× CPU and 1× RAM
- total ≈ $7.20/mo
- modeled: Official US East (N. Virginia) Graviton rates and separate image/state meters; AWS labels capacity in GB, compared 1:1 to the shared GiB workload, CPU and memory burst are independently supplied, aggregate hours use equal-duration inference to enforce the eight-hour total-lifecycle and per-suspension limits, and data transfer is excluded.

## Ranked default prices

| rank | product_id | product | gross_usd_month | modeled_invoice_usd_month | invoice_context | vs_lambda_gross | billed allocation | evidence | first-party pricing source |
| ---: | --- | --- | ---: | ---: | --- | ---: | --- | --- | --- |
| 1 | koyeb | Koyeb Sandboxes | 1.44 | 1.44 | Starter | 0.20008x | 0.5 vCPU; 0.5 GiB RAM; 5 GiB disk; selection: micro | modeled | [source](https://www.koyeb.com/docs/reference/instances) |
| 2 | fly-machines | Fly Machines | 1.807969621 | 1.807969621 | same as gross | 0.251208x | 1 vCPU; 0.5 GiB RAM; 5 GiB disk; 6 GB retained-storage billing; selection: shared-cpu-1x / 512 MB | modeled | [source](https://fly.io/docs/about/pricing/) |
| 3 | sandbox0 | Sandbox0 | 2.1 | 2.1 | same as gross | 0.291784x | 0.5 vCPU; 0.5 GiB RAM; 8 GiB disk; selection: documented custom template | modeled | [source](https://sandbox0.ai/pricing) |
| 4 | sailboxes | Sailboxes | 2.175 | 2.175 | same as gross | 0.302204x | 0.25 vCPU; 0.5 GiB RAM; 5 GiB disk; selection: S | modeled | [source](https://docs.sailresearch.com/sailboxes-pricing) |
| 5 | upstash | Upstash Sandbox | 3.036870912 | 3.036870912 | Pay as you go | 0.421957x | 0.25 vCPU; 0.5 GiB RAM; 5 GiB disk; 5.36870912 GB retained-storage billing; selection: Medium | modeled | [source](https://upstash.com/pricing/box) |
| 6 | aws-agentcore | AWS AgentCore Code Interpreter | 3.1825 | 3.1825 | same as gross | 0.442191x | 0.25 vCPU; 0.5 GiB RAM; 5 GiB disk | modeled | [source](https://aws.amazon.com/bedrock/agentcore/pricing/) |
| 7 | exe-dev | exe.dev | 3.25 | 20 | Personal pool | 0.45157x | 0.25 vCPU; 0.5 GiB RAM; 5 GiB disk | modeled | [source](https://exe.dev/sandbox) |
| 8 | box-by-ascii | Box by ASCII | 3.6 | 20 | monthly minimum applied | 0.5002x | 2 vCPU; 4 GiB RAM; 12 GiB disk; selection: small | modeled | [source](https://docs.ascii.dev/box/billing) |
| 9 | tensorlake | Tensorlake | 3.659589041 | 3.659589041 | Usage Credits | 0.50848x | 0.25 vCPU; 0.5 GiB RAM; 10 GiB disk | modeled | [source](https://www.tensorlake.ai/pricing) |
| 10 | northflank | Northflank Sandboxes | 4.083 | 4.083 | same as gross | 0.567311x | 0.5 vCPU; 1 GiB RAM; 5 GiB disk; selection: nf-compute-50 | modeled | [source](https://northflank.com/pricing) |
| 11 | blaxel | Blaxel | 4.85739726 | 4.85739726 | same as gross | 0.674909x | 0.25 vCPU; 0.5 GiB RAM; 5 GiB disk | modeled | [source](https://blaxel.ai/pricing) |
| 12 | langsmith | LangSmith Sandboxes | 5.17395 | 0 | recurring account allowance applied | 0.718892x | 0.25 vCPU; 0.5 GiB RAM; 5 GiB disk | modeled | [source](https://www.langchain.com/pricing) |
| 13 | azure-container-apps | Azure Container Apps | 5.4 | 0 | recurring account allowance applied | 0.750301x | 0.25 vCPU; 0.5 GiB RAM; 20 GiB disk; selection: XS | modeled | [source](https://azure.microsoft.com/en-us/pricing/details/container-apps/) |
| 14 | modal | Modal Sandboxes | 5.949 | 0 | recurring account allowance applied | 0.826581x | 0.25 vCPU; 0.5 GiB RAM; 5 GiB disk | modeled | [source](https://modal.com/pricing) |
| 15 | aws-lambda-microvms | AWS Lambda MicroVMs | 7.197115111 | 7.197115111 | same as gross | 1x | 0.25 vCPU; 0.5 GiB RAM; 5 GiB disk; selection: micro-0.25 | modeled | [source](https://aws.amazon.com/lambda/pricing/) |
| 16 | fly-io-sprites | Fly.io Sprites | 7.286803403 | 7.286803403 | same as gross | 1.012462x | 0.25 vCPU; 0.5 GiB RAM; 5 GiB disk; 0.536870912 GB metered RAM; 5.36870912 GB hot-storage billing; 5.36870912 GB cold-storage billing | modeled | [source](https://fly.io/sprites/) |
| 17 | cloudflare | Cloudflare Sandboxes | 7.47058294 | 11.74518294 | required plan fee applied | 1.037997x | 1 vCPU; 3 GiB RAM; 5 GiB disk; 5.36870912 GB hot-storage billing; selection: custom | modeled | [source](https://developers.cloudflare.com/sandbox/platform/pricing/) |
| 18 | isorun | Isorun | 8 | 8 | same as gross | 1.111556x | 1 vCPU; 1 GiB RAM; 5 GiB disk | modeled | [source](https://docs.isorun.ai/getting-started/pricing) |
| 19 | novita | Novita AI Sandbox | 8.464087425 | 8.208 | recurring account allowance applied | 1.176039x | 1 vCPU; 0.5 GiB RAM; 5 GiB disk; 5.36870912 GB cold-storage billing; selection: 1 vCPU | modeled | [source](https://novita.ai/docs/guides/sandbox-pricing) |
| 20 | opencomputer | OpenComputer | 9.48 | 9.48 | Pay as you go | 1.317194x | 0.25 vCPU; 0.5 GiB RAM; 8 GiB disk; selection: 0.25 vCPU / 0.5 GB | modeled | [source](https://opencomputer.dev/) |
| 21 | freestyle | Freestyle | 9.6679 | 0 | Free | 1.343302x | 1 vCPU; 0.5 GiB RAM; 5 GiB disk | modeled | [source](https://www.freestyle.sh/docs/vms/pricing-and-limits) |
| 22 | morph | Morph Cloud | 10.015676631 | 10.015676631 | same as gross | 1.391624x | 0.25 vCPU; 0.5 GiB RAM; 5 GiB disk; 2.147483648 GB retained-storage billing | modeled | [source](https://cloud.morph.so/web/subscribe) |
| 23 | beam | Beam Cloud | 11.52 | 0 | recurring account allowance applied | 1.600641x | 0.25 vCPU; 0.5 GiB RAM; 5 GiB disk | modeled | [source](https://docs.beam.cloud/v2/resources/pricing-and-billing) |
| 24 | e2b | E2B | 11.7 | 161.7 | Pro | 1.625651x | 1 vCPU; 0.5 GiB RAM; 5 GiB disk; selection: 1 vCPU | modeled | [source](https://e2b.dev/pricing) |
| 25 | superserve | Superserve | 11.7 | 11.7 | same as gross | 1.625651x | 1 vCPU; 0.5 GiB RAM; 5 GiB disk | modeled | [source](https://api.superserve.ai/billing/pricing/public) |
| 26 | vercel | Vercel Sandbox | 11.840036 | 20 | Pro | 1.645109x | 1 vCPU; 2 GiB RAM; 64 GiB disk; selection: 1 vCPU | modeled | [source](https://vercel.com/docs/sandbox/pricing) |
| 27 | declaw | Declaw | 12.31056 | 100 | Pro | 1.710485x | 1 vCPU; 0.5 GiB RAM; 20 GiB disk | modeled | [source](https://docs.declaw.ai/platform/billing) |
| 28 | tenki | Tenki Sandbox | 12.42504 | 2.42504 | Starter | 1.726392x | 1 vCPU; 0.5 GiB RAM; 5 GiB disk | modeled | [source](https://tenki.cloud/pricing) |
| 29 | instavm | InstaVM | 12.43 | 11.7 | Free | 1.727081x | 1 vCPU; 0.5 GiB RAM; 5 GiB disk | modeled | [source](https://instavm.io/pricing) |
| 30 | createos | CreateOS Sandbox | 13.32 | 13.32 | same as gross | 1.850742x | 1 vCPU; 1 GiB RAM; 5 GiB disk; selection: s-1vcpu-1gb | verified | [source](https://createos.sh/products/sandbox) |
| 31 | daytona | Daytona | 13.32 | 13.32 | same as gross | 1.850742x | 1 vCPU; 1 GiB RAM; 5 GiB disk | modeled | [source](https://www.daytona.io/pricing) |
| 32 | cloud-run | Cloud Run Sandboxes (Jobs) | 13.899178 | 8.859178 | recurring account allowance applied | 1.931215x | 1 vCPU; 0.5 GiB RAM; 10 GiB disk; selection: 1 vCPU | modeled | [source](https://cloud.google.com/run/pricing) |
| 33 | miosa | MIOSA | 14.70438 | 14.70438 | Developer | 2.043094x | 1 vCPU; 1 GiB RAM; 20 GiB disk; selection: Tiny | modeled | [source](https://miosa.ai/pricing) |
| 34 | runloop | Runloop | 17.3395368 | 265.84 | Pro | 2.409234x | 0.5 vCPU; 1 GiB RAM; 6 GiB disk | modeled | [source](https://runloop.ai/pricing) |
| 35 | islo | Islo | 18.7 | 18.7 | same as gross | 2.598263x | 1 vCPU; 0.5 GiB RAM; 5 GiB disk | modeled | [source](https://islo.dev/pricing/) |
| 36 | codesandbox | CodeSandbox SDK | 30 | 176 | Scale | 4.168337x | 2 vCPU; 4 GiB RAM; 20 GiB disk; selection: Nano | modeled | [source](https://codesandbox.io/pricing) |
| 37 | archil | Archil | 55.5 | 53.865 | recurring account allowance applied | 7.711423x | 2 vCPU; 4 GiB RAM; 5 GiB disk; selection: Developer sandbox | modeled | [source](https://archil.com/pricing) |

## Scope notes for ranked prices

- **Koyeb Sandboxes** (modeled): Official Sandbox, instance, billing, plan, lifecycle, and Cloud Hypervisor sources; the Sandbox-specific $0.0036/hour advertised floor supports the Standard schedule instead of the cheaper Eco schedule, current docs elsewhere define 1 GB of RAM as 2^30 bytes, account-wide paid-plan compute allowances may be consumed elsewhere, and temporary no-cost Light Sleep is excluded. [First-party pricing source](https://www.koyeb.com/docs/reference/instances).
- **Fly Machines** (modeled): Official exact per-second Ashburn preset, volume, stopped-rootfs, and snapshot rates; GiB quantities convert by bytes, volumes round up to Fly’s integer decimal-GB size, stored image bytes represent stopped rootfs, optional automatic daily volume snapshots are disabled because their incremental stored bytes are not an input, and network usage is excluded. [First-party pricing source](https://fly.io/docs/about/pricing/).
- **Sandbox0** (modeled): Official configured-memory and actual retained-byte rates plus a template schema that explicitly documents a reproducible 500m CPU / 512Mi memory / 8Gi ephemeral-storage configuration; this is a conservative documented configuration rather than a claimed minimum, retained bytes use the explicit cold-disk input, and currently free network plus requests are excluded. [First-party pricing source](https://sandbox0.ai/pricing).
- **Sailboxes** (modeled): Official observed-usage and size-specific creation rates plus S/M/L capacity ceilings; the default selects S, observed CPU uses the workload utilization input, observed RAM and NVMe disk use explicit workload inputs, provider sampling occurs about every 15 seconds, and inactive lifecycle phases plus network are excluded. [First-party pricing source](https://docs.sailresearch.com/sailboxes-pricing).
- **Upstash Sandbox** (modeled): Official active-core, actual-storage, and box-size rules; binary GiB inputs convert byte-exactly to decimal GB, so the 5 GiB default requires Medium, the Free plan is enforced as independent CPU and literal GB hard caps, and optional fixed keep-alive pricing is excluded. [First-party pricing source](https://upstash.com/pricing/box).
- **AWS AgentCore Code Interpreter** (modeled): Official active-CPU and per-second cumulative-peak-memory rates, one-second minimum, 128 MB memory billing floor, and session resource ceilings; the observed-RAM input is the time-weighted average of that per-second meter including system overhead, AWS labels memory GB and it is compared 1:1 to the shared GiB workload, new boxes stand in for sessions, the dedicated-microVM documentation does not identify the VMM, and persistent S3 Files/EFS plus EC2 network charges are excluded. [First-party pricing source](https://aws.amazon.com/bedrock/agentcore/pricing/).
- **exe.dev** (modeled): Official GiB-based usage and Personal-pool rates; gross mode is marginal usage, invoice mode includes the required $20 active subscription, usage CPU and RAM are billed by hourly peak approximated from aggregate utilization and metered RAM, and inactive-memory storage plus transfer are excluded. [First-party pricing source](https://exe.dev/sandbox).
- **Box by ASCII** (modeled): Official current size rates, monthly plan floor, and guaranteed data floors; provider-published GB is compared 1:1 to the shared GiB workload, and current docs identify a Linux VM without naming its infrastructure host or VMM. [First-party pricing source](https://docs.ascii.dev/box/billing).
- **Tensorlake** (modeled): Official Free, Usage Credits, and Pro rates and limits, including Pro’s $250 recurring usage allowance; 2-hour and 24-hour per-session caps use equal-duration inference from aggregate hours and new sandboxes, snapshot storage is prorated over explicit suspended hours, active CPU can fall back to allocated CPU when a sample is rejected, the prepaid opening balance and concurrency are unknown, and GB is compared 1:1 to GiB. [First-party pricing source](https://www.tensorlake.ai/pricing).
- **Northflank Sandboxes** (modeled): Official deployment plans and per-second usage rates; the requested Lambda-sized box rounds to nf-compute-50, the API represents 1024 MB as one GiB while the price page labels the unit GB, so that configured allocation is priced 1:1, and the non-production Developer free tier plus egress are excluded. [First-party pricing source](https://northflank.com/pricing).
- **Blaxel** (modeled): Official allocated-memory compute, provisioned-volume, standby-snapshot, and image-storage rates; the independent disk request is supplied by a volume because the writable root layer is an approximately half-RAM tmpfs, snapshot size and suspended duration are explicit, image size is user-supplied, and all listed GB units are provider-defined GiB. The public API exposes arbitrary MB memory but does not publish a minimum or step, so the requested 512 MiB is a modeled allocation. [First-party pricing source](https://blaxel.ai/pricing).
- **LangSmith Sandboxes** (modeled): Official LCU/LSU-derived meters and current 0.05-to-16-vCPU, millicore, and 2-to-6-GiB-per-vCPU constraints; the Lambda request fits at 250 millicores and 0.5 GiB, retained filesystem capacity is modeled across the sandbox lifetime because stop/start preserves it, recurring allowances are account-wide, and Sandboxes became generally available in May 2026. [First-party pricing source](https://www.langchain.com/pricing).
- **Azure Container Apps** (modeled): Official Sandbox tiers and Consumption reference rates; tier capacity is published in GB and compared 1:1 to GiB, Azure prices vary by region/currency, and request and persistent-volume charges are excluded. [First-party pricing source](https://azure.microsoft.com/en-us/pricing/details/container-apps/).
- **Modal Sandboxes** (modeled): Official base-region meters, 0.125-physical-core floor represented as 0.25 vCPU, default 512 GiB ephemeral quota, and recurring Starter credit; selected regions add 15 to 75 percent and the account-wide credit may be consumed elsewhere. [First-party pricing source](https://modal.com/pricing).
- **AWS Lambda MicroVMs** (modeled): Official US East (N. Virginia) Graviton rates and separate image/state meters; AWS labels capacity in GB, compared 1:1 to the shared GiB workload, CPU and memory burst are independently supplied, aggregate hours use equal-duration inference to enforce the eight-hour total-lifecycle and per-suspension limits, and data transfer is excluded. [First-party pricing source](https://aws.amazon.com/lambda/pricing/).
- **Fly.io Sprites** (modeled): Official actual-usage rates, 0.0625-vCPU and 256-MB running floors, and 8-vCPU / 16-GB / 100-GB capacity bounds; sampled memory and hot/cold blocks come from explicit workload inputs and convert byte-exactly from GiB to decimal GB, while aggregate hours cannot reproduce the provider’s sampling timeline or split storage transitions over time. [First-party pricing source](https://fly.io/sprites/).
- **Cloudflare Sandboxes** (modeled): Official container meters, predefined and custom instance constraints, and byte-exact GiB-to-GB disk conversion; the lowest raw-meter valid configuration is selected for the supplied utilization, while Workers, Durable Objects, logs, egress, and image storage are excluded because the workload has no corresponding inputs. [First-party pricing source](https://developers.cloudflare.com/sandbox/platform/pricing/).
- **Isorun** (modeled): Official per-second allocated-vCPU and RAM rates and zero hibernation charge; the SDK exposes numeric resource fields but publishes 1 vCPU / 1 GiB as its smallest documented shape, so smaller requests conservatively round up, configurable scratch disk has no listed meter, and promotional credit is excluded. [First-party pricing source](https://docs.isorun.ai/getting-started/pricing).
- **Novita AI Sandbox** (modeled): Official rates and paid-tier limits; paused storage and its 60 GB account allowance use byte-exact decimal-GB quantities, the retained-disk input is treated as this workload’s stored data, and aggregate hours cannot prove the longest individual session. [First-party pricing source](https://novita.ai/docs/guides/sandbox-pricing).
- **OpenComputer** (modeled): Official current baseline and peak rates; the burst slider is a time-weighted average because exact peak seconds are not an input, the published per-minute figures are rounded, and one-time signup credit is excluded. [First-party pricing source](https://opencomputer.dev/).
- **Freestyle** (modeled): Official provisioned-resource meters, 1-vCPU published floor, paid retained disk while paused, plan limits, and recurring allowances; the smallest base image is BusyBox rather than a general agent image, and allowances are account-wide and may already be consumed by other VMs. [First-party pricing source](https://www.freestyle.sh/docs/vms/pricing-and-limits).
- **Morph Cloud** (modeled): Official MCU compute and snapshot-capacity rules; snapshot size is user-supplied and converts byte-exactly from GiB to the published decimal-GB storage meter, a 730-hour month converts snapshot MCU-hours, and starting plan-credit renewal cadence is not published. [First-party pricing source](https://cloud.morph.so/web/subscribe).
- **Beam Cloud** (modeled): Official CPU-core and memory rates plus recurring Developer credit; the SDK sends fractional CPU as millicores and exposes Gi memory, while the rate card labels memory GB. Beam defines neither guest-thread topology nor a byte conversion, so CPU and memory units are mapped 1:1 with both ambiguities disclosed, and the credit is account-wide. [First-party pricing source](https://docs.beam.cloud/v2/resources/pricing-and-billing).
- **E2B** (modeled): Official resource rates and plan limits; invoice mode uses the implied equal session duration to choose Hobby or Pro, aggregate inputs cannot prove the longest individual session or concurrency, and the explicit Hobby / Pro rate table is used despite broader Pro-only customization copy. [First-party pricing source](https://e2b.dev/pricing).
- **Superserve** (modeled): The live first-party billing endpoint marks CPU and memory billable at the displayed rates and storage non-billable; the static price page still says paused storage is charged, so the current total follows the live billing flag and discloses that conflict. [First-party pricing source](https://api.superserve.ai/billing/pricing/public).
- **Vercel Sandbox** (modeled): Official meters, persistence behavior, and plan rules; provider-published GB is compared 1:1 to GiB, the per-session memory minimum and plan eligibility use new boxes plus resumes with an implied equal duration, while regional multipliers, request traffic, concurrency, and account-wide usage are not derivable from aggregate inputs. [First-party pricing source](https://vercel.com/docs/sandbox/pricing).
- **Declaw** (modeled): Official provisioned-resource meters, template resource ranges, fixed 20 GB overlay, and plan rules; published GB and API MB are compared 1:1 to GiB because byte semantics are undefined, aggregate hours imply an average session, the one-time grant is excluded, and invoice mode applies Pro's monthly recharge commitment because the fixed disk exceeds Free. [First-party pricing source](https://docs.declaw.ai/platform/billing).
- **Tenki Sandbox** (modeled): Official component rates, 1-core minimum, disk and storage allowance distinction, resource ceilings, and recurring plan credits; GB is compared 1:1 to GiB, the Team total uses month-to-month pricing because commitment is not an input, and account-wide credits may be consumed elsewhere. [First-party pricing source](https://tenki.cloud/pricing).
- **InstaVM** (modeled): Official rates, integer 1-vCPU floor, and billing mechanics; the volume API explicitly defines the published 10 GB quota as 10 GiB, gross mode excludes its account-wide allowance, invoice mode applies it and selects the cheapest eligible plan, and temporary signup credit is excluded. [First-party pricing source](https://instavm.io/pricing).
- **CreateOS Sandbox** (verified): Rates and represented billing mechanics were checked against the linked first-party source. [First-party pricing source](https://createos.sh/products/sandbox).
- **Daytona** (modeled): Official per-second reserved-resource rates, lifecycle billing, and public organization limits of 4 vCPU, 8 GiB RAM, and 10 GiB disk; enterprise limits may differ, the workload represents one retained disk allocation, and temporary trial credit is excluded. [First-party pricing source](https://www.daytona.io/pricing).
- **Cloud Run Sandboxes (Jobs)** (modeled): Official Cloud Run Sandbox, us-central1 Jobs, ephemeral-disk, resource-limit, per-task minimum, seven-day cap, and free-tier sources; one host task allocation is attributed to one sandbox, aggregate host billable time is spread equally across tasks, the nested isolation mechanism is undisclosed, actual task overhead and retries are not inputs, and account-wide free usage may be consumed elsewhere. [First-party pricing source](https://cloud.google.com/run/pricing).
- **MIOSA** (modeled): Official resource rates, pause lifecycle, plan minimums, and current public Sandbox shape contracts; published GB/MB are compared 1:1 to GiB, selected shape storage is charged at running and cold rates across a 730-hour month, transient fleet-readiness state is not treated as a pricing constraint, one-time credit and egress are excluded. [First-party pricing source](https://miosa.ai/pricing).
- **Runloop** (modeled): Official resource and storage rates plus current SDK constraints; the shared request rounds to the valid custom 0.5-vCPU, 1-GiB, 6-GiB allocation, provider pricing labels memory and storage as GB while its resource API labels them GiB, aggregate inputs infer equal session duration for the 48-hour lifetime cap, transition time is excluded, suspend/resume requires Pro, and account-wide storage allowance usage is unknown. [First-party pricing source](https://runloop.ai/pricing).
- **Islo** (modeled): The live canonical rate card bills provisioned CPU, provisioned memory, and disk only while running; an older first-party page described actual cgroup usage, so the current live source governs. The API accepts only integer vCPU requests, making 1 vCPU the smallest representable allocation that satisfies the 0.25-vCPU request, but this is not asserted as the provider default. GB is compared 1:1 to GiB because bytes are undefined, and the implementation remains unknown because the dedicated-microVM claim does not identify a VMM. [First-party pricing source](https://islo.dev/pricing/).
- **CodeSandbox SDK** (modeled): Official current Nano template, $0.15 on-demand VM-hour rate, $0.015 credit price, plan allowances, and resource ceilings; only Nano is modeled because larger tier credit multipliers are unpublished, Build included hours are enforced as a hard cap, Scale is evaluated in invoice mode, and the plan's 20 GB disk is compared 1:1 to GiB because byte semantics are undefined. [First-party pricing source](https://codesandbox.io/pricing).
- **Archil** (modeled): Official US East Developer allowance and overage rates; the capacity input is not the billed storage basis, so actual retained bytes come from the explicit cold-disk input, GB is compared 1:1 to GiB, storage-linked free compute is excluded because the current rate card does not define its interaction with plan allowances, the VMM is not disclosed, and Team has different account-wide rates but cannot be cheaper within this calculator’s single-box workload range. [First-party pricing source](https://archil.com/pricing).

## Self-hosted and infrastructure reference prices

AWS EC2 `t4g.nano` is the permanent raw-VM baseline and remains visible in the default human calculator. The other rows are hidden by the calculator's default “show other self-host” setting. None receive a managed-product rank. They can still be useful as infrastructure baselines, but may require a cluster, VM, or other capacity that is shared outside one sandbox.

| product_id | product | gross_usd_month | modeled_invoice_usd_month | invoice_context | vs_lambda_gross | billed allocation | evidence | first-party pricing source |
| --- | --- | ---: | ---: | --- | ---: | --- | --- | --- |
| gke-agent-sandbox | GKE Agent Sandbox | 2.85615 | 2.85615 | same as gross | 0.396847x | 0.25 vCPU; 0.5 GiB RAM; 5 GiB disk | modeled | [source](https://cloud.google.com/kubernetes-engine/pricing) |
| ec2-t4g-nano | AWS EC2 t4g.nano | 3.466 | 3.466 | same as gross | 0.481582x | 2 vCPU; 0.5 GiB RAM; 5 GiB disk; selection: t4g.nano | modeled | [source](https://aws.amazon.com/ec2/pricing/on-demand/) |
| ec2-t4g-small | AWS EC2 t4g.small | 12.664 | 12.664 | same as gross | 1.759594x | 2 vCPU; 2 GiB RAM; 5 GiB disk; selection: t4g.small | modeled | [source](https://aws.amazon.com/ec2/pricing/on-demand/) |
| ec2-c7g-medium | AWS EC2 c7g.medium | 26.899 | 26.899 | same as gross | 3.73747x | 1 vCPU; 2 GiB RAM; 5 GiB disk; selection: c7g.medium | modeled | [source](https://aws.amazon.com/ec2/pricing/on-demand/) |
| ec2-c7g-large | AWS EC2 c7g.large | 53.325 | 53.325 | same as gross | 7.409219x | 2 vCPU; 4 GiB RAM; 5 GiB disk; selection: c7g.large | modeled | [source](https://aws.amazon.com/ec2/pricing/on-demand/) |

### Scope notes for reference prices

- **GKE Agent Sandbox** (modeled): Official no-surcharge statement, us-central1 Autopilot requested-resource rates, current bursting-cluster constraints, and mandatory managed gVisor configuration; the numeric result is a single running Pod’s marginal cost and excludes shared cluster fee/credit, warm-pool Pods, persistent storage, Pod snapshots, network, and logging. [First-party pricing source](https://cloud.google.com/kubernetes-engine/pricing).
- **AWS EC2 t4g.nano** (modeled): Permanent default-visible raw-VM baseline using official US East (N. Virginia) Linux on-demand instance, T4 CPU-credit, Unlimited-mode, and gp3 storage rates; burst overage uses an aggregate monthly approximation because the workload does not contain AWS's rolling 24-hour CPU timeline, and network, OS overhead, administration, and taxes are excluded. [First-party pricing source](https://aws.amazon.com/ec2/pricing/on-demand/).
- **AWS EC2 t4g.small** (modeled): Official US East (N. Virginia) Linux on-demand instance, T4 CPU-credit, Unlimited-mode, and gp3 storage rates; burst overage uses an aggregate monthly approximation because the workload does not contain AWS's rolling 24-hour CPU timeline, the temporary t4g.small trial is excluded, and network, OS overhead, administration, and taxes are excluded. [First-party pricing source](https://aws.amazon.com/ec2/pricing/on-demand/).
- **AWS EC2 c7g.medium** (modeled): Official US East (N. Virginia) Linux on-demand instance and gp3 storage rates; network, OS overhead, administration, and taxes are excluded. [First-party pricing source](https://aws.amazon.com/ec2/pricing/on-demand/).
- **AWS EC2 c7g.large** (modeled): Official US East (N. Virginia) Linux on-demand instance and gp3 storage rates; network, OS overhead, administration, and taxes are excluded. [First-party pricing source](https://aws.amazon.com/ec2/pricing/on-demand/).

## No default price

| product_id | product | state | reason | source when available |
| --- | --- | --- | --- | --- |
| deno-sandbox | Deno Sandbox | unsupported-default | 3.333h implied average session exceeds 0.5h published limit | [source](https://deno.com/deploy/pricing) |
| ellipsis | Ellipsis | unsupported-default | 3.333h implied average session exceeds 1h published limit | [source](https://www.ellipsis.dev/pricing) |
| scrapybara | Scrapybara | unverified | The current first-party site says Scrapybara built computers for agents and is now building Capy. It no longer offers or prices the former Scrapybara service publicly, so historical plans are removed and this row remains unranked. | [source](https://scrapybara.ai/) |
| railway | Railway Sandboxes | unverified | Railway publishes per-minute VM beta CPU and RAM rates and says sandboxes use them while running. However, its current official SDK exposes no CPU or RAM sizing option or resource fields, and its product docs do not publish the sandbox capacity, billing floor, or whether guest overhead is included. The exact Lambda-sized workload therefore cannot be validated from first-party public evidence, so this row is unranked. Hobby and Pro usage-minimum rules remain informational only. | [source](https://docs.railway.com/pricing/plans) |
| computer-agents | Computer Agents | unverified | The exact current per-minute compute tiers are public; the displayed hourly figures are rounded, so the stored rates use per-minute price times 60. The pricing page does not disclose durable storage capacity or an overage rate for this retained-disk workload. A separate feature page mentions 4 GB without defining whether it is per computer, persistent, included, or billable, so this row remains unranked rather than omitting storage. | [source](https://computer-agents.com/pricing) |
| arker | Arker | unverified | Current first-party deployment documentation says there is no free tier and access is being prioritized for paying customers, while the public pricing route requires authentication and publishes no rate card. Invite-only access is not a reproducible list price, so the row remains unranked. | [source](https://arker.ai/pricing) |
| mosaic | Mosaic | unverified | The current first-party Mosaic site identifies the sandbox product, resolving the former benchmark-only vendor identity. It publishes no public pricing or billing rate, so the row remains unranked. | [source](https://www.mosaicos.com) |
| run-cloud | Run Cloud | unverified | Per-second billing; one physical core equals two vCPUs. CPU size is a billing floor and usage above it auto-scales at the same rate. The public pricing page lists CPU and memory, while the product page explicitly says storage is also metered. Because no storage rate is published, a complete 5 GiB sandbox total cannot be reproduced and this row is unranked. The advertised $15 monthly credit is also excluded. | [source](https://run.cloud/pricing/) |
| sandbox-as-a-service | Sandbox as a Service | unverified | The current Sandy by Chutes page replaced the old per-second tiers with $17/$29/$59 monthly plans and says customers pay only for usage, but it does not publish included sandbox-hours, concurrency, or the exact relationship between plan price and usage. The stale $0.09/$0.28/$0.55 hourly table is removed and this row is unranked. | [source](https://sandbox-as-a-service.com/pricing) |
| cleanroom | Cleanroom | self-hosted | No managed-service list price applies. | n/a |
| smolvm-smol-machines | smolvm (smol-machines) | self-hosted | No managed-service list price applies. | n/a |
| leap0 | Leap0 | unverified | Current preview is free without a card, not a durable $0 rate for comparison. Published post-preview CPU/RAM rates are not charged yet. | [source](https://leap0.dev/) |
| mitos | Mitos | self-hosted | No managed-service list price applies. | n/a |
| microsandbox | microsandbox | self-hosted | No managed-service list price applies. | n/a |
| omnirun | OmniRun | unverified | The default CPU and RAM fit OmniRun’s published standard sandbox, so an unpublished larger-configuration multiplier is not the blocker. The current page publishes EUR 0/19/79 plans, 25/100/600 included standard sandbox-hours, and EUR 0.19/0.15 overages. However, neither the pricing page nor public docs publish the sandbox disk capacity needed to prove the 5 GiB request fits, and the site compares totals in USD without a time-stamped exchange-rate input. OmniRun also documents ephemeral sessions rather than retained state. The row therefore remains unranked for explicit workload and currency gaps. | [source](https://omnirun.io/pricing) |
| boxlite | BoxLite | self-hosted | No managed-service list price applies. | n/a |
| qbox | Qbox | self-hosted | No managed-service list price applies. | n/a |
| katakate | Katakate (k7) | self-hosted | No managed-service list price applies. | n/a |
| agentenv | AgentENV | self-hosted | No managed-service list price applies. | n/a |
| arrakis | Arrakis | self-hosted | No managed-service list price applies. | n/a |
| smolvm-celesto | SmolVM (Celesto AI) | self-hosted | No managed-service list price applies. | n/a |
| baponi | Baponi | unsupported-default | 3.333h implied average execution exceeds 1h published limit | [source](https://baponi.ai/pricing) |
| opensandbox | OpenSandbox (Alibaba) | self-hosted | No managed-service list price applies. | n/a |
| aio-sandbox | AIO Sandbox | self-hosted | No managed-service list price applies. | n/a |
| kubernetes-agent-sandbox | Kubernetes Agent Sandbox (kubernetes-sigs) | self-hosted | No managed-service list price applies. | n/a |
| openhands | OpenHands (sandboxed runtime) | self-hosted | No managed-service list price applies. | n/a |
| deepinfra | DeepInfra Sandboxes | unverified | Launched 2026-08. The public quote is nano from $0.054/hr, but per-plan vCPU, RAM, and disk are only exposed via the authenticated GET /v1/sandboxes/catalog endpoint, so a fair TCO cannot be modeled yet. | [source](https://deepinfra.com/blog/sandboxes) |
| coreweave-sandboxes | CoreWeave Sandboxes | unverified | Current first-party product and documentation pages confirm two different commercial paths. CKS sandboxes schedule onto the customer’s existing contracted CoreWeave cluster capacity, so a sandbox-only marginal price depends on that customer’s infrastructure contract and utilization. Serverless sandboxes are managed through Weights & Biases, but neither company publishes a separate sandbox resource rate that maps this workload. The row remains unranked rather than assigning generic CoreWeave node pricing to an unknown shared cluster. | [source](https://www.coreweave.com/products/coreweave-sandboxes) |

## Interpretation and exclusions

- `verified` means the represented rates and mechanics are directly supported by current first-party evidence.
- `modeled` means first-party rates are used, but the result depends on a disclosed assumption shown above.
- Rankings and `vs_lambda_gross` always use gross usage. Invoice totals are additional context, not a second ranking, because account-wide allowances may already be consumed elsewhere.
- `unverified` and `pricing-unavailable` products never receive a numeric rank.
- `unsupported-default` means this precise workload fails a published resource, lifecycle, execution, or plan constraint.
- Totals exclude taxes and, unless the provider note says otherwise, egress, request traffic, regional multipliers, concurrency effects, negotiated discounts, model-token charges, and other usage in the same account.
- GiB and GB are not silently interchangeable when the provider exposes byte-defined billing or configuration units. Where semantics are not published, the assumption is stated in that provider's scope note.
- Display formatting does not round the internal calculation. Numeric table values retain up to nine decimal places.

For the full source-linked product objects use <https://api.sandboxprices.com/api/products>. For custom calculations use <https://api.sandboxprices.com/api/tco> or the interactive calculator.