What AI Sandboxes Cost After the Session Ends
Compare six AI sandbox billing models and calculate how snapshots, short sessions, storage and plan limits change a monthly bill.
An AI code sandbox can keep costing money after the code stops. The running machine, a saved snapshot, an attached drive, data transfer and the plan can appear as separate charges. On Vercel Pro, snapshot storage is listed at $0.08 per GB-month across regions. Drive storage is $0.05 per GB-month in the default iad1 region and varies by region, alongside active CPU and provisioned memory charges. Blaxel lists standby snapshots at $0.20 per GB-month. Those are published rates, not a prediction of your bill: you need the actual stored sizes and retention times to calculate it. Vercel Sandbox pricing and Blaxel pricing show the relevant meters.
The useful comparison is a month of sandbox lifecycles, from creation through running, pausing, saving state and expiration. An hourly compute quote alone misses what remains after a session ends. The six services below use different units and products, so compare each bill on its own terms.
Compare what each sandbox meters
These are public USD list rates read on November 8, 2026. A vCPU is not a Modal physical core, and a vendor’s GB may not be another vendor’s GiB. The table does not normalize hardware or claim equal performance. Follow the maker’s page before using a figure in a purchase decision.
| Sandbox | Running compute meter | State after execution | Plan or limit to check |
|---|---|---|---|
| E2B | $0.000014 per vCPU-second plus $0.0000045 per GiB-second of RAM while running | Pricing table includes 10 GiB storage on Hobby or 20 GiB on Pro; a separate saved-snapshot rate is not published there | Hobby: $100 one-time credit, one-hour sessions and 20 concurrent sandboxes. Pro: $150/month plus usage, 24-hour sessions and 100 concurrent sandboxes |
| Daytona | $0.0504 per vCPU-hour plus $0.0162 per GiB-hour of memory | Storage: $0.000108 per GiB-hour after the first 5 GiB | The pricing page lists $200 in free compute. Check account limits and storage duration for your workload |
| Modal Sandboxes | $0.00003942 per physical-core-second plus $0.00000667 per GiB-second; billed on the higher of requested or used resources | This Sandbox rate does not state a universal snapshot storage price; check the storage product you actually use | Starter: $0 plus compute, with $30 monthly included compute. Team: $250 plus compute |
| Vercel Sandbox | In default iad1 region, $0.128 per active CPU-hour plus $0.0212 per provisioned GB-hour of memory | Snapshot: $0.08 per GB-month across regions. Drive: $0.05 per GB-month in iad1; rates vary by region, with separate read/write meters. | Hobby quotas include 5 active CPU-hours and 420 GB-hours of memory per month; Pro usage draws on a $20 monthly credit |
| Blaxel Sandboxes | $0.0000115 per allocated GB RAM-second while active; CPU scales with memory | Standby snapshots: $0.20 per GB-month. Provisioned volumes: $0.12 per GB-month | Pay-as-you-go lists up to $200 free credits; concurrent sandbox quotas rise by tier |
| Cloudflare Sandbox on Containers | On Workers Paid, overage is $0.000020 per active vCPU-second plus $0.0000025 per provisioned GiB-second while the container runs | Provisioned disk overage: $0.00000007 per GB-second while running. Persistent storage needs a separate product and price | Workers Paid is $5/month. The Worker and Durable Object can also be billed |
E2B’s price page bills seconds of a running sandbox and lists included storage capacity, but it does not price every possible retained-state configuration. Do not turn “included storage” into a promise that all snapshots, mounts or long-term files are free. Daytona publishes storage separately; its free 5 GiB allowance matters when many workspaces persist. Modal explicitly prices Sandboxes at its non-preemptible rates, which are three times its general list compute prices, and charges for the higher of requested resources or actual use. E2B, Daytona and Modal document those distinctions.
Vercel separates active CPU from provisioned memory. Its pricing page also meters creations, network, snapshot storage, drive storage, reads and writes. Blaxel moves idle sandboxes to standby and charges for the saved memory and filesystem snapshot while compute is suspended. Cloudflare’s Sandbox runs on Containers: CPU is active-use based, but memory and disk are based on the provisioned instance while it runs; Workers and Durable Objects have their own meters. Vercel, Blaxel and Cloudflare publish these rules. None tells you the size of your particular saved workspace.
Count running work and retained state separately
Make one row for each sandbox or workspace, then one row for each retained snapshot or drive. Record the provider’s billing unit as printed. This avoids counting a stopped machine as running compute or overlooking files that remain after it stops.
| Field to record | Why it changes the calculation |
|---|---|
| Workspace ID; creation, start, stop and resume times | Separates wall-clock existence from metered running time and counts creations |
| Allocated CPU and memory; active CPU seconds; billed memory time | Captures providers that bill CPU activity, provisioned memory, or both |
| Snapshot ID, creation and expiration time, billed stored GB | Turns retained state into GB-days or the provider’s stated monthly unit |
| Attached drive or volume, provisioned size, read and write GB | Separates storage capacity from operations and transferred data |
| Network GB, plan fee, credits and quota tier | Keeps costs and allowances that do not appear in the compute rate |
Use the provider’s usage export or invoice to fill the billed size. A 2 GB allocated disk is not proof that a snapshot bills 2 GB; compression, shared blocks and metadata can change stored bytes. A pause may end compute charging on one service while leaving a storage charge. Saving files to a drive is a different action again. Check what the API actually stopped, what it preserved, and which meter continued.
For a monthly estimate, calculate running CPU and memory according to their own rules, then add each storage object’s billed size multiplied by its retained time. Add drives, I/O, network and any required plan fee. Subtract only credits your account can actually use in that billing period. Your model API calls are another line: the site’s AI API price comparison and API cost calculator help with tokens, not sandbox infrastructure. For worker time between inference requests, see serverless GPU pricing; that is a different product from these code-execution workspaces.
Price a month of saved workspaces
Here is an illustrative scenario, not measured usage or a vendor invoice. Suppose a team creates 100 new snapshots at the start of each day. Assume each snapshot’s billed stored size is 2 GB, the month has 30 days, and there is no starting stock. A snapshot occupies storage through its creation day and expires before new snapshots arrive after its chosen 1-, 7- or 30-day retention period. For this example only, divide GB-days by 30 to express GB-months. Vercel publishes a $0.08 per GB-month Pro snapshot rate, but its page does not specify that this 30-day normalization exactly matches invoice proration. The dollar column is therefore an illustration from that rate, before credits or other charges. Vercel’s pricing and expiration rules say snapshots expire 30 days after last use by default, with configurable expiration.
| Retention of each new snapshot | Stored GB at end of day 30 | First-month GB-days | Example GB-months (GB-days ÷ 30) | Example storage amount at $0.08/GB-month |
|---|---|---|---|---|
| 1 day | 200 | 6,000 | 200 | $16.00 |
| 7 days | 1,400 | 37,800 | 1,260 | $100.80 |
| 30 days | 6,000 | 93,000 | 3,100 | $248.00 |
The seven-day row ramps from 200 GB on day one to 1,400 GB on day seven, then holds that stock for the remaining 23 days. Its total is 200 × (1 + 2 + 3 + 4 + 5 + 6) + 1,400 × 24 = 37,800 GB-days. The 30-day row is 200 × (1 + … + 30) = 93,000 GB-days. A mature 30-day cycle with 30 cohorts already in place would hold 6,000 GB every day, or 6,000 GB-months over a 30-day example month; it is much higher than the first month’s ramp. A mature seven-day cycle holds 1,400 GB, equivalent to $112 in this simplified month.
Consider a different policy: keep only one latest snapshot for each of 100 continuing workspaces. If each latest snapshot really bills 2 GB and replacement never overlaps, the stock is 200 GB, or $16 across an illustrative 30-day month. That comparison changes the number of retained objects, not their measured size. Real systems may preserve earlier blocks or briefly overlap replacement snapshots. Measure the usage export before treating the difference as a saving. Choose a retention period that meets recovery and compliance needs, with a tested backup policy before removing any production state.
Short sessions can hit a billing floor
Vercel says provisioned memory is billed in one-minute minimum increments. It does not say on that page that every additional partial minute is rounded up in the same way, so this example covers only sub-minute runs. Suppose there are 1,000 separate ten-second sandboxes, each provisioned with 4 GB of memory in the default iad1 region. Seconds-only arithmetic would be 1,000 × 4 GB × 10/3,600 hours × $0.0212 = $0.2356. With at least one minute of memory per run, it becomes 1,000 × 4 GB × 60/3,600 hours × $0.0212 = $1.4133, before rounding on an invoice. That is a minimum-memory illustration, not a whole-sandbox bill. CPU activity, creations and other meters remain separate; Hobby’s included quota or Pro credit may change the cash amount. Vercel documents the memory minimum and quotas.
Do not apply that memory rule to E2B, Daytona, Modal, Blaxel or Cloudflare without their own billing documentation. Nor should you assume that bundling unrelated users into one long-running sandbox is an acceptable cost fix: isolation and data boundaries are part of the product choice.
Check the plan before the hourly price
An apparently cheap rate may require a tier your workload cannot avoid. E2B Hobby is capped at 20 concurrent sandboxes and one-hour sessions, while Pro lists $150 per month, 100 concurrent sandboxes and 24-hour sessions. Vercel Hobby lists 10 concurrent sandboxes and a 45-minute maximum, while Pro lists 10,000 concurrent sandboxes and 24 hours. Cloudflare’s Containers require Workers Paid, and Blaxel’s concurrency increases with top-up tiers. Compare these published limits with your observed peak concurrency and longest real task, not an average. E2B, Vercel, Cloudflare and Blaxel list the relevant plan terms.
A one-time onboarding credit is not a repeatable monthly allowance. E2B calls its $100 Hobby credit one-time; Modal lists $30 monthly included compute on Starter; Vercel Pro applies its $20 monthly credit to Sandbox usage. Keep those in separate columns in your budget. For each candidate, calculate a month with no promotional credit, then compare it with the first month. If a storage price, retention rule or quota is not published for your chosen setup, mark that line unknown and ask the provider or inspect a small billed pilot. No service is universally cheapest from these rate cards alone.
Sources
- E2B pricing: running CPU and RAM, included storage and plan limits.
- Daytona pricing: CPU, memory and storage rates.
- Modal pricing: Sandbox-specific compute rates, resource basis, plans and credits.
- Vercel Sandbox pricing and quotas: eight usage meters, minimum memory billing, snapshot expiration, plans and credits.
- Blaxel pricing: active compute, standby snapshots, volumes and quotas.
- Cloudflare Sandbox overview and Containers pricing: underlying meters and Workers Paid terms.
All pages read on November 8, 2026.
Free toolWhat will 10,000 chats a month cost?Estimate my bill
