| Provider | Billing model | $/month |
|---|
The billing model matters more than the rate
E2B and Daytona charge the identical $0.0504/vCPU-hour and $0.0162/GB-hour whether the sandbox is running a test suite or sitting idle between agent steps — both are "wall-clock" billers. Modal bills wall-clock too, at a higher per-unit rate. Vercel Sandbox and Cloudflare Sandbox only charge their full CPU rate while code is actively executing. Fly.io Sprites go further: billing pauses to a near-zero storage-only rate roughly 30 seconds after activity stops. For a typical coding agent — which spends most of a session waiting on model output or user input, not executing — that distinction swings the bill far more than any vendor's headline per-hour price.
Who actually wins depends on your busy %
At high utilization (an agent that's running code 80-100% of its open session — batch processing, CI-style workloads), the wall-clock providers' lower per-unit rates can come out ahead since there's little idle time to waste. At low utilization (a chat-style coding assistant where the sandbox mostly waits on the user or the LLM, often 5-25% busy), active-only and usage-metered providers pull dramatically ahead because they simply don't bill the idle majority of the session. There's no universal winner — run your own numbers above, not someone else's benchmark scenario, since the "best" sandbox is a function of your agent's actual usage pattern, not a fixed ranking.
Tools & Hosting
How this calculator works
Monthly cost = (vCPU × $/vCPU-hour + RAM × $/GB-hour) × hours/month × sandboxes, except each provider's rate applies to a different slice of time depending on its billing model: wall-clock providers (E2B, Daytona, Modal) bill the full open-session hours regardless of your busy %; active-only providers (Vercel, Cloudflare) apply your busy % to the CPU term only, since memory stays provisioned the whole session; Fly.io Sprites apply busy % to both CPU and memory, since its usage metering pauses billing on both once activity stops. This is why the ranking changes as you move the busy % slider — it isn't one number being scaled, it's different providers responding to idle time in structurally different ways.
📚 New here? Learn how to choose an LLM →
FAQ
Why do sandbox providers cost so differently for the same vCPU and RAM?
The billing model matters more than the headline rate. E2B, Daytona and Modal bill wall-clock time — once a sandbox starts, you pay for every second it's reserved, busy or not. Vercel and Cloudflare only charge full CPU rate while code is actively running. Fly.io Sprites pause billing (down to a tiny storage-only rate) about 30 seconds after activity stops. For a coding agent that spends most of its session waiting on the model or the user, that distinction can be a bigger cost lever than the per-unit price.
What does "busy %" mean in this calculator?
The share of a sandbox's open session where code is actually executing, as opposed to sitting idle waiting for the next agent step. A chat-style coding agent that runs a test suite every few minutes might be 5-15% busy; a sandbox doing continuous batch processing can be 80-100% busy. Wall-clock providers charge the same either way; active-only and usage-metered providers scale down with it.
Are these the exact current prices?
Rates are the published per-vCPU-hour and per-GB-hour figures as of the verified date below, cross-checked across each vendor's pricing page. Plan minimums, free credits and volume discounts aren't modeled — this estimates the raw usage-based rate, which is what dominates once you're past the free tier.