Vercel Sandbox's $0.128 Active-CPU Hour vs E2B's $0.0504: Where Idle-Free Billing Wins — and Where Hetzner Wins by 28x

Vercel charges $0.128 per active CPU-hour for sandboxes and nothing while they idle. E2B charges $0.0504 per vCPU-hour, flat, while the sandbox runs. Read those two numbers once and Vercel looks 2.5x more expensive — until your agent spends most of its loop waiting on the model, at which point Vercel looks like a bargain. Both readings are wrong in the interesting way: the headline rates are not comparable at all, because one meters active CPU and the other meters wall-clock runtime, and the memory line underneath them never sleeps on either platform.

This post normalizes the rate cards to one comparable unit — a typical 2 vCPU / 4 GB agent sandbox over a 730-hour month — and finds the exact crossover: below about 32% CPU-active time, Vercel's idle-free billing wins; above it, E2B's flat rate wins by a widening margin. And once sandboxes stop idling much at all, both meters lose to bin-packing onto owned hardware by more than an order of magnitude.

The verdict in one table​

Here is the monthly cost of one 2 vCPU / 4 GB sandbox that exists for the full month, at three CPU-active fractions. Vercel and Fly Sprites scale with active time; E2B and Modal are shown always-running (their idle story is pausing or killing the sandbox, covered below); Hetzner is one eighth of a dedicated AX52 (Ryzen 7 7700, 16 threads, 64 GB, €64/month).

Three things jump out. First, the Vercel-vs-E2B crossover sits at roughly 32% CPU-active: an agent that thinks (model round-trips, human-in-the-loop waits) more than two-thirds of the time is cheaper on Vercel; an agent that compiles, tests, or scrapes continuously is cheaper on E2B. Second, Fly Sprites undercuts every metered vendor at low utilization because both its CPU and memory lines scale with active use — but its memory rate is the steepest in the industry, so it converges toward Vercel at full burn. Third, owned hardware is not in the same league on unit price: about $9 per sandbox-month against $121–$249 on the meters. The rest of this post substantiates every cell and then asks when each row actually applies.

The rate cards, normalized to one comparable unit​

Every figure below comes from the vendor's own pricing page or a documented 2026 comparison, and all of them describe the same thing: what a 2 vCPU / 4 GB sandbox costs per hour it exists.

Vercel Sandbox (pricing docs, updated September 2026) splits billing in two: CPU meters at $0.128 per active CPU-hour — zero while the sandbox idles — while memory meters at $0.0212 per GB-hour provisioned, i.e. wall-clock for as long as the sandbox exists. Data transfer is $0.15/GB beyond a 20 GB Hobby allotment, creations are a negligible $0.60 per million, and snapshot storage is $0.08/GB-month. Our 4 GB sandbox therefore carries a $61.90/month memory floor before it does a single flop of work.

E2B and Daytona have converged on identical list rates (Northflank's 2026 comparison): $0.0504/vCPU-hour plus $0.0162/GiB-hour, billed per second while the sandbox runs. A 2 vCPU / 4 GB sandbox costs about $0.166/hour running, or $120.89 for a full month that never pauses. E2B's idle story is lifecycle, not metering: kill or auto-pause the sandbox and the meter stops entirely — memory included, unlike Vercel.

Modal sandboxes bill per second at $0.1419 per physical-core-hour (one core equals 2 vCPUs) plus $0.0242/GiB-hour — roughly $0.239/hour, or $174.25/month always running. Modal's real advantage is GPUs (L4 to H200 on one bill), which none of Vercel, E2B, or Fly Sprites offer on sandboxes at all; on CPU-only math it is the priciest row in the table.

Fly.io Sprites (pricing) charges $0.07 per CPU-hour measured from actual cgroup usage plus $0.04375/GB-hour of memory, and a Sprite that goes idle stops accruing both — you keep paying only storage (hot NVMe at $0.00068/GB-hour). That auto-sleep is why Sprites wins the 10% column so decisively at $23.00. The catch is the memory rate: at full burn, 4 GB costs $127.75/month in memory alone, more than double Vercel's memory floor.

Hetzner is the anchor: an AX52 dedicated box (8 cores / 16 threads, 64 GB RAM, 2×1 TB NVMe) rents for €64/month. Bin-pack eight 2 vCPU / 4 GB sandboxes onto it — CPU-bound at eight, memory-bound at sixteen — and each sandbox-month costs €8, roughly $9. No per-second meter, no plan floor, no transfer bill; Hetzner dedicated bandwidth is effectively unmetered.

Active-CPU is worth ~2x on idle loops — and ~0x on hot loops​

The crossover math is simple enough to do on a napkin. Vercel's monthly bill for our sandbox shape is $186.88 × u + $61.90, where u is the CPU-active fraction; E2B's always-running bill is a flat $120.89. Set them equal and u ≈ 0.32. An August 2026 independent benchmark (MarkTechPost) lands in the same place empirically, finding active-CPU billing "worth roughly 2x on idle-heavy loops" — our table shows 1.5x at 10% active once Vercel's memory floor is included.

Which side of 32% is your agent on? It depends on what the sandbox is waiting for:

• Below the line (Vercel wins): interactive coding agents where each tool call is followed by a multi-second model round-trip, human-approval gates, or long-polling waits. The sandbox CPU is genuinely idle most of the wall-clock hour, and Vercel meters exactly that.
• Above the line (E2B wins): agents running test suites, builds, benchmarks, scrapers, or data pipelines — workloads where the sandbox CPU is hot for minutes at a stretch. At 100% active, Vercel costs 2.06x E2B ($248.78 vs $120.89), and the gap is pure CPU rate: $0.128 vs $0.0504.

Two honest caveats keep this from being a Vercel-vs-E2B commercial. First, E2B can auto-pause an idle sandbox to zero — including memory — which beats Vercel's $61.90 memory floor for workloads with long idle gaps if you accept pause/resume latency and lifecycle complexity. Vercel's idle discount is automatic; E2B's is architectural. Second, Vercel's memory line never sleeps: a sandbox that exists but does nothing still costs $0.0848/hour per 4 GB. Teams that leave hundreds of sandboxes provisioned "just in case" pay the floor on all of them.

The lines nobody puts in the headline​

Headline rates decide the table above; the lines below decide whether the table applies to you at all.

Plan floors. E2B's Hobby tier is free with $100 in one-time credits but caps you at 20 concurrent sandboxes and 1-hour sessions; exceeding either means Pro at $150/month plus usage (Beam's 2026 guide), which buys 24-hour sessions and 100 concurrent sandboxes. Vercel flips the shape: a $20/month Pro credit absorbs the first chunk of metered usage, so small experiments often bill $0 above the plan. At fleet scale both floors amortize to noise; at prototype scale the $150 E2B floor is the single biggest line on the invoice.

Concurrency and session caps. Vercel allows 10 concurrent sandboxes on Hobby and 10,000 on Pro/Enterprise, with max session durations of 45 minutes (Hobby) and 24 hours (Pro/Enterprise). E2B Pro tops out at 100 concurrent sandboxes and 24-hour sessions. If your design needs 500 always-on sandboxes, E2B's per-sandbox math is irrelevant until you've had the enterprise conversation; if you need week-long sessions, neither vendor's standard tier offers them.

Transfer and storage. Vercel's $0.15/GB transfer matches hyperscaler egress and bites agents that download datasets or container layers repeatedly — the 20 GB Hobby allotment evaporates fast. E2B includes 10–20 GB of storage free per sandbox; Vercel charges $0.08/GB-month for snapshots plus $0.004/GB of Drive writes. None of these dominate a code-execution workload, but a data-heavy agent can easily spend more on transfer than on CPU.

The 200-sandbox fleet math​

Now scale the table to a realistic always-on fleet: 200 concurrent 2 vCPU / 4 GB sandboxes running hot (u ≈ 100%) for a full month.

Owned hardware is roughly 14x cheaper than E2B and 28x cheaper than Vercel at full utilization — and the gap widens with scale because the metered side is perfectly linear while the owned side steps in €64 increments with bin-packing slack you can fill with more tenants. This is the same crossover the industry keeps rediscovering: per-second billing wins when utilization is low and spiky; flat capacity wins when utilization is high and predictable. "Zero charge when idle" is a genuine innovation in metering — it just stops mattering the moment your sandboxes aren't idle much.

The honest counter-argument is everything the €1,600 excludes: the Firecracker/Kata control plane, image distribution, autoscaling, network policy, observability, and the on-call rotation for all of it. E2B's own enterprise pitch concedes the microVM is the easy 5%. A team pricing the Hetzner row must add at least one platform engineer's attention — but at a $22,000+/month delta to the cheapest metered option, that engineer pays for themselves many times over once the fleet is steady-state.

Pick by workload shape, not by vendor​

The rate cards above compress into four decision rules:

• Bursty, interactive agents with long think-waits → Vercel Sandbox or Fly Sprites. Automatic idle discounting, no pause/resume architecture to build, and Vercel's $20 Pro credit covers experimentation. Watch the memory floor and the 24-hour session cap.
• Sustained batch-style agent work (test suites, builds, evaluations) → E2B or Daytona at $0.0504/vCPU-hour, ideally with auto-pause for the gaps. Budget the $150/month Pro floor from day one if you need more than 20 concurrent sandboxes.
• GPU-bearing agent workloads → Modal (or Northflank's $0.01667/vCPU-hour BYOC-adjacent pricing if you want one bill for CPU and GPUs). The CPU-sandbox vendors above simply do not offer GPUs.
• A steady-state fleet of always-on sandboxes → owned Hetzner capacity with an open-source isolation stack (Firecracker, Kata, gVisor). The moment average utilization crosses ~50% across hundreds of sandboxes, every per-second meter becomes a tax on predictability.

The deeper lesson is that 2026's sandbox pricing war is really a metering war: active-CPU, per-second-while-running, and sleep-when-idle are three different answers to "what counts as using a computer." Each one is cheapest exactly for the workload shape it was designed around — and each one's marketing reads as universal right up until you multiply by 730 hours and 200 sandboxes. Do that multiplication before you pick, not after the first invoice.

Running agent sandboxes on machines you own is exactly the workload a self-hosted PaaS should make boring: bin-pack Firecracker microVMs onto flat Hetzner capacity instead of renting them back per second. Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own. Star the repo on GitHub or deploy your first app today.