Sandbox costs
28 September, 2026
One always-on box per user.
2 GB RAM · 30 days / 720 hours · 1 vCPU where configurable. No sleep savings.
| Provider | CPU / memory | Per box / month | Account adjustment | Always-on fit | |
|---|---|---|---|---|---|
| Archil | 2 vCPU / 4 GB | $194.40 | 30 min/account included | 4 GB minimum · lifetime unverified | |
The calculation 720 × $0.27 = $194.40 per box. Account total for N boxes: $194.40 × N − $0.135 included time; divide by N. One box: $194.27. The time credit is shared. What this buys 2 vCPU · 4 GB · 10 GB performance storage Published sandbox size is 4 GB, not 2 GB. Developer plan includes 10 GB performance storage. Customer bucket / archive storage and out-of-region transfer are additional. No month-long session guarantee established. | |||||
| AWS AgentCore v1 | 1 vCPU / 2 GB | $13.61–78.05⁺ | No fixed plan modeled | Usage only · 8-hour session cap | |
The calculation 2 GB × 720 × $0.00945 = $13.608 memory, plus $64.44 × average vCPU usage. With CPU averaging 0–1 vCPU: $13.61–78.05, before extras. This is a usage model, not a reserved 2-GB box. What this buys 1 vCPU consumed · 2 GB used Usage equivalent only: microVM lifetime is capped at 8 hours. Requires replacement environments; not one uninterrupted monthly machine. ECR/S3 artifacts, logs and other AWS services are separate. Rates shown for the published US example. The pricing page separately lists v2 consumption at $0.1276/vCPU-hour and $0.0169/GB-hour: $116.21 for the same modeled usage. Future committed-baseline discounts are excluded. | |||||
| Blaxel | Bundled CPU / 2 GB | $60.82⁺ | No fixed plan modeled | Keep-alive + eligible TTL tier | |
The calculation 2 × 2,592,000 × $0.0000115 = $59.62 compute; 10 GB × $0.12 = $1.20 volume. What this buys 2 GB · CPU scales with RAM · 10 GB persistent volume CPU is bundled into allocated RAM. For 720 awake hours, compute is $59.616 and 10 GB volume is $1.20. Add image storage at $0.045/GB-month. Blaxel documents process keep-alive or connection activity to prevent standby; use a quota tier allowing unlimited persistence and disable expiration. Starter TTL restrictions do not establish month-long runtime. | |||||
| Boat / formerly Box | 2 vCPU / 4 GB | $12.96 | $20/month account minimum | 4 GB minimum · override TTL | |
The calculation 720 × $0.018 = $12.96 per Small box. Account total = max($20, $12.96 × N). Per user = that total ÷ N. One user: $20; two users: $12.96 each. What this buys Small: 2 vCPU · 4 GB Box now redirects to Boat. Small is the nearest published size above 2 GB. The minimum is per account, not per VM. Bundled filesystem; no extra disk modeled. Default 1-hour TTL must be overridden; maximum TTL not verified. | |||||
| Cloudflare Sandbox | ½ vCPU / 4 GiB | $27.37–53.29⁺ | $5 plan + shared allowances | 4 GiB minimum · ½ vCPU | |
The calculation Memory: 4 × 720 × $0.009 = $25.92. Disk: 8 × 720 × $0.000252 = $1.45152. CPU: $51.84 × average vCPU usage (0–0.5). Per box: $27.37–53.29 before shared adjustments. What this buys standard-1: ½ vCPU · 4 GiB · 8 GB ephemeral disk No exact 2-GiB shape. This is the smallest listed size with at least 2 GiB; the earlier table unnecessarily required 1 vCPU and chose 6 GiB. Add one $5 account plan, then subtract available allowances (25 GiB-hours, 375 CPU-minutes, 200 disk GB-hours) once. One box: $32.10–57.57 before Workers, Durable Objects, logs, traffic and durable storage. If you need 1 full vCPU, standard-2 is $41.06–92.90 per box before the account adjustment. CPU range assumes zero to full CPU consumption while awake; memory and disk are allocated charges. | |||||
| CreateOS | 1 vCPU / 2 GiB | $43.87 | No minimum advertised | Disable auto-pause | |
The calculation 720 × ($0.03616363 + 2 × $0.01159025 + 10 × $0.0001584) = $43.87. What this buys 1 vCPU · 2 GiB · 10 GB disk 720 allocated hours, even if the guest is idle. Egress is unmetered; the pricing page advertises no minimum or monthly commitment. Auto-pause is off by default and there is no maximum session length. The product FAQ still lists plan-based concurrency caps, so confirm fleet quotas before scaling. Signup credits excluded. | |||||
| Daytona | 1 vCPU / 2 GiB | $60.00 | Signup credit excluded | Disable auto-stop / TTL | |
The calculation 720 × ($0.0504 + 2 × $0.0162) = $59.62 compute; 5 billable GiB × 720 × $0.000108 = $0.39 disk. What this buys 1 vCPU · 2 GiB · 10 GiB disk First 5 GiB disk included; total uses unrounded components ($60.0048). Disable auto-stop, auto-pause where applicable, and wall-clock TTL. Signup credits excluded. Container/VM capabilities differ: verify the chosen sandbox class. | |||||
| Declaw | 1 vCPU / 2 GB | $60.72 | $100 monthly account deposit | Usage only · 72-hour cap | |
The calculation 2,592,000 × ($0.000014 + 2 × $0.0000045 + 10 × $0.0000000424) = $60.72. What this buys 1 vCPU · 2 GB · 10 GB overlay disk Free has a 1-hour session cap; Pro has 72 hours and requires $100 deposited each calendar month. Unspent deposits carry forward: $100 cash funding is not an extra $100 fee. Neither proves uninterrupted 30-day runtime. Guardrails scans cost extra. | |||||
| E2B | 1 vCPU / 2 GiB | $59.62 | +$150/month Pro account | Usage only · 24-hour cap | |
The calculation 2,592,000 × ($0.000014 CPU + $0.000009 memory) = $59.616 per box. Pro account total = $150 + $59.616 × N. Per user = $59.616 + $150/N. One user: $209.62; 100 users: $61.12 each. What this buys Pro custom: 1 vCPU · 2 GiB · 20 GiB included disk The $150 is one account fee, not a fee for every end user. Pro allows custom resources and sessions up to 24 hours; 720 hours requires replacement sessions. Thus this is not one uninterrupted month-long box. Enterprise persistent sessions require a quote; concurrency limits still apply. | |||||
| exe.dev | 2 vCPU / 8 GB pool | Not per-box | $20 Personal / $25 Team seat | 8 GB / 2 vCPU shared pool | |
The calculation Personal: $20/month for a shared 2-vCPU / 8-GB account pool. Team: $25 per provider seat/month with its own 2-vCPU / 8-GB pool. Neither quotes a marginal 2-GB box price. What this buys Personal pool: 2 vCPU · 8 GB · 100 GB pooled used disk One user’s workload may fit in the $20 Personal pool. Dividing $20 by four because 8 GB holds four 2-GB workloads would leave only 0.5 vCPU each at full contention and assumes permitted app-user hosting. Provider seats are not automatically your application users. Confirm commercial hosting terms and capacity before using this for per-user economics. | |||||
| Freestyle | 1 vCPU / 2 GiB | $49.59 | $50 min + shared allowances | 32 GB disk modeled | |
The calculation Per box: 720 × ($0.04032 + 2 × $0.0129 + 32 × $0.000086) = $49.58784 before account allowances. For N boxes: compute = max(720N − 200, 0) × $0.04032 + max(1440N − 400, 0) × $0.0129; disk = max(23040N − 60000, 0) × $0.000086. Account total = max(plan minimum, compute + disk); divide by N. What this buys 1 vCPU · 2 GiB · 32 GB disk modeled Allocated CPU and RAM are charged while the VM runs, even when idle. One box is $50 on Hobby after shared allowances. Hobby supports 40 concurrent VMs; Pro has a $500 minimum and 400 concurrent VMs. Shared allowances apply once. At 100 users on Pro, this model gives $49.40/user. A smaller valid disk can lower the per-box price; 32 GB is the modeled disk, not a universal minimum. | |||||
| Google Cloud Run | 1 vCPU / 2 GiB | $57.02⁺ | Shared free tier available | Container service; instances replaced | |
The calculation 2,592,000 × ($0.000018 + 2 × $0.000002) = $57.024 per instance. With unused account free tier, subtract $5.22 once per account: one instance is $51.80; 100 instances average $56.97 each. What this buys 1 vCPU · 2 GiB · Instance-based · us-central1 Instance-based billing in us-central1 with minimum instances maintained. CPU allocation is billed while idle too. The filesystem uses RAM; external durable storage is additional. Cloud Run can replace instances: this buys continuously provisioned service capacity, not one permanent VM. Free tier is shared across the billing account. | |||||
| Isorun | 1 vCPU / 2 GiB | $39.80 | No minimum advertised | Disable auto-destroy | |
The calculation 720 × ($0.025 + 2 × $0.015) = $39.60 compute; 10 GB × $0.02/month = $0.20 storage. Rates ↗ · Disk ↗ · Lifetime ↗ What this buys 1 vCPU · 2 GiB · 10 GB actual persistent data Uses published per-resource rates; request this custom size. Persistent disk bills actual data, not provisioned capacity. No minimum or egress fees. Disable the auto-destroy timeout (0) and do not hibernate. | |||||
| Modal Sandbox | 1 vCPU / 2 GiB | $85.67 | $30/account recurring credit | Usage only · 24-hour cap | |
The calculation 2,592,000 × (0.5 physical core × $0.00003942 + 2 GiB × $0.00000667) = $85.6656 per box. Starter account total = max(0, $85.6656 × N − $30). One user: $55.67; 100 users: $85.37 each, subject to account concurrency. What this buys 0.5 physical core (= 1 vCPU) · 2 GiB · no extra volume With 1 vCPU reserved, idle time still bills the reservation: max(requested, actual). Use cpu=0.5 physical cores and memory=2048 MiB; cap bursts for this model. Sandbox-specific rates, not cheaper Function rates. The recurring $30 credit is shared. Team fees and credits differ. The 24-hour session cap requires replacements; this is not one uninterrupted monthly box. | |||||
| Northflank | 1 vCPU / 2 GB | $25.50 | No fixed plan modeled | Container service · 10 GB disk | |
The calculation Published $24/month compute plan + 10 × $0.15/GB-month disk = $25.50. What this buys nf-compute-100-2 · 1 dedicated vCPU · 2048 MB · 10 GB SSD Uses the listed monthly plan price rather than multiplying rounded hourly numbers. Generic service plan; dedicated sandbox-isolation/custom deployments may need a separate quote. Egress costs $0.06/GB. | |||||
| OpenComputer | 1 vCPU / 2 GB | $136.08 | $20 Pro → $200 usage credit | Agent sessions · not verified 24/7 | |
The calculation 43,200 minutes × $0.00315 = $136.08 of metered machine usage. Pro account: $20 includes $200 prepaid usage; modeled bill = $20 + max(0, $136.08 × N − $200). One user: $20; two: $46.08/user. Max: $200 includes $2,000 usage; at 10 users this is $20/user if all credits apply to machine time. What this buys Default: 1 vCPU · 2 GB · Current agent-session offering These are shared plan credits, not an automatic $20 price for each application user. No model spend or automatic size burst is modeled. The public offer describes agent sessions; standalone always-awake sandbox access, concurrency and no-sleep controls remain unverified. Do not use this as a confirmed 24/7 infrastructure quote. | |||||
| Render | 1 CPU / 2 GB | $27.50 | Hobby workspace: $0 | Container service · 10 GB disk | |
What this buys 1c-2g: 1 CPU · 2 GB · 10 GB persistent SSD Free Hobby workspace plus paid service compute. This is a web service/background worker, not a hardware-isolated general-purpose VM. Includes workspace allowances; bandwidth, builds and higher workspace plans can add charges. | |||||
| Sprites | 1 CPU / 2 GB used | $67.92–118.32⁺ | Optional prepaid bundles | 2 GB used · keep-awake required | |
The calculation 2 GB RAM × 720 × $0.04375 + 10 GB hot disk × 720 × $0.000683 = $67.9176. Add $50.40 × average CPU usage (0–1 CPU): $67.92–118.32 per box, before bundles and extra storage. What this buys 1 CPU fully used · 2 GB used · 10 GB hot storage Two GB of actual resident memory is modeled, not merely an available 2-GB limit. Idle CPU can be cheaper; RAM usage and CPU bursting can change the bill. Keep awake with documented activity or renewable task holds. Hero is $100/account with 1,200 CPU-hours, 4,800 RAM GB-hours and 150 GB storage; one fully busy modeled box fits for $100, but those allowances do not repeat per user. Cold/checkpoint storage is extra. | |||||
Your cost per user = total provider bill ÷ number of users. Each user gets one box. Multiply per-box usage by your user count, apply account fees, minimums and credits once, then divide. Expand a row for formulas and one-user examples. The main price excludes shared account adjustments.
Awake is not busy. Ranges show approximately idle through fully used CPU, with the stated RAM held throughout. Allocation-billed services charge the same while awake and idle. CPU is 1 vCPU where configurable; larger or shared shapes are explicit. GB/GiB follow each provider’s units.
Disk and runtime. Includes 10 GB disk where separately modeled; bundled disk, Freestyle’s 32 GB, and exceptions are in the details. A + means additional costs remain. Session-capped rows price 720 aggregate hours, not one box that survives a month. Taxes, models, traffic, backups and application overhead are excluded. Use 744 hours for a 31-day month; subscriptions stay monthly.
A few that need qualification
These providers publish a plan or usage price, but it doesn’t establish the same always-on machine.
| Provider | Published price | What remains unverified | Evidence |
|---|---|---|---|
| Unikraft Cloud | $0 Hobby plan 2 running instances, 2 total vCPU, 4 GiB total RAM | The current page advertises enough resources for the example, but does not clearly establish a 720-hour compute allowance or all storage/usage charges. Treat $0 as the advertised plan fee, not a verified free always-on invoice. | Official pricing |
| Kernel (hosted) | $172.80 usage / shared $5 credit 4-GB app invocation equivalent | 2,592,000 seconds × $0.0000666668/sec, with a $5 Developer credit applied once per account. This is a hosted invocation meter, not a quoted 2-GB always-on VM. Browser usage is billed separately. Hypeman below is a different, self-hosted product. | Official rates & allocation |
Self-hosted tools and infrastructure
These do not have a comparable hosted 2-GB sandbox price. You must buy the host first; dividing its RAM by 2 GB does not guarantee enough CPU, disk, or isolation overhead for that many fully busy guests.
- SlicerVM + Latitude.sh: orchestration software plus a separately billed host. Total = the chosen Latitude server + Slicer licence + storage/network costs.
- Hypeman: self-hosted VM manager, not Kernel hosted browser pricing. Total depends on your physical or nested-virtualization host.
- celld: local runtime software; host infrastructure is separate.
- AWS EC2, Azure VMSS, Google Compute Engine, Hetzner, Latitude, Scaleway and Vultr: these are underlying host infrastructure options. A 2-GB VM quote would price a different product from the larger KVM-capable hosts used in those studies. Host region, SKU, CPU class, disk and guest density must be specified to calculate cost per sandbox.
Checked against official pricing pages. Calculations use unrounded rates; totals are rounded to cents. Account allowances apply once, not once per VM. This estimates infrastructure usage; commercial quotas and continuous-runtime eligibility still apply.