Dora NodaE2B will spin up an isolated Linux microVM for your coding agent in about 150 milliseconds and bill you for every second it lives. A single Hetzner box you own will run the same Firecracker isolation all month for roughly the price of one E2B Pro subscription — before E2B meters a single sandbox-second. Both statements are true at the same time, and which one matters depends on a single variable most teams never compute: how many sandbox-hours you actually burn.
This post does the worked comparison the TODO promised: what a team running moderate agent-sandbox volume really pays across E2B's tiers against the marginal cost of running the same Firecracker isolation on already-provisioned Hetzner capacity. The short version, delivered before the math: under ~400 sandbox-hours a month, E2B's Hobby tier is nearly free and the right answer. Past that, a flat owned box wins on dollars — and for GPU-backed agent work, E2B isn't even in the race, because its sandboxes are CPU-only at every tier.
The answer up front: three teams, two bills
Assumptions, stated once so every table below is checkable. The reference sandbox is E2B's default: 2 vCPU and 4 GiB RAM. E2B's published 2026 rates are $0.000014 per vCPU-second ($0.0504/vCPU-hour) and $0.0000045 per GiB-second ($0.0162/GiB-hour), so the default box costs $0.1656 per running hour. The owned-hardware side is one Hetzner AX52 — 8 cores, 64 GB RAM, 2 TB NVMe — at roughly €64 (~$70) a month flat. Firecracker's overhead is about 5 MB of memory per microVM with ~125 ms boots, so a 64 GB box comfortably hosts dozens of 2 vCPU / 4 GiB sandboxes the way E2B packs its own hosts.
| Team | Monthly sandbox-hours | E2B bill | Owned AX52 bill | Delta |
|---|---|---|---|---|
| Weekend tinkerer: 1 sandbox, ~20 hrs/month | 20 | ~$3.31 (Hobby, $0 plan fee) | $70 flat ($0 marginal if the box already exists) | E2B wins by ~$67 |
| 5-person agent team: 10 concurrent sandboxes, 8 hrs/day, 22 workdays | 1,760 | ~$291 usage + $150 Pro = ~$441 | $70 flat | Owned wins by ~$371 |
| Heavy burst shop: 50 concurrent average, around the clock | 36,500 | ~$6,044 usage + $1,150 Pro++ = ~$7,194 | ~$210 (three boxes) | Owned wins by ~$6,980 |
Read the shape of the table, not just the rows. At low volume the meter is nearly free and the flat box looks wasteful; at steady volume the flat box wins by an order of magnitude. The crossover for this sandbox shape sits around 400 sandbox-hours a month — roughly one default sandbox running half the day, every day. Most teams cross it the week their second agent workflow goes to production without noticing. And the third row carries the post's second punchline: E2B's meter has no GPU line item at any price, so the "heavy" team doing inference-adjacent agent work can't buy its way out on E2B at all.
What E2B actually charges — and what the 150 milliseconds buy
The meter: plan fee plus provisioned-seconds. E2B's pricing is two lines. The plan fee buys concurrency and session length: Hobby is $0 with 20 concurrent sandboxes capped at 1-hour sessions and a one-time $100 usage credit; Pro is $150/month with 100 concurrent sandboxes, 24-hour sessions, and custom CPU/RAM shapes; concurrency add-ons step up to 600 sandboxes for $650 (Pro+) and 1,100 for $1,150 (Pro++). Usage on top is wall-clock provisioned capacity — every second a sandbox exists with N vCPU and M GiB, you owe N × $0.000014 plus M × $0.0000045. There is no active-CPU distinction: a sandbox idling while the agent waits on a model response bills at full price, the same trap as Railway-style resource-second meters. Storage is included (10 GiB Hobby, 20 GiB Pro), which is the one line you never have to think about.
The speed: Firecracker's physics, productized. The sub-500 ms headline is real and it is also mostly not E2B's invention. Firecracker boots a stripped guest kernel in ~125 ms with ~5 MB of VMM overhead per microVM, and snapshot resume cuts that further — E2B, Daytona, and Fly Sprites all land in the 90–200 ms band because they all sit on the same primitive. What E2B productizes is the combination: start times that fast times per-second billing means spinning a fresh kernel per task costs nothing in either latency or money, so "one sandbox per agent step" becomes a sane architecture instead of an extravagance. That pairing — not either half alone — is what you are renting.
The ceiling: CPU-only, stated plainly on the pricing page. "Sandboxes are CPU-only; E2B does not offer GPU sandboxes." Rivals made the opposite call: Modal and Beam put A100/H100-class GPUs behind their sandbox APIs, Daytona ships GPU options, and Fly's Sprites price CPU and memory with accelerator-adjacent persistence. If your agent sandboxes only run code, linters, and browsers, the ceiling never touches you. The moment a sandbox needs to touch a model — local inference, embedding batches, GPU-dependent test suites — E2B exits the comparison and the rest of this post's math stops applying.
The worked math, step by step
The tinkerer: the meter at its most flattering. Twenty sandbox-hours at $0.1656/hr is $3.31, plan fee $0, concurrency 1 against a limit of 20, sessions under an hour. The $100 one-time credit covers the first two and a half years at this pace. The owned box at $70/month is pure waste here — unless it already exists for another reason, in which case its marginal sandbox cost is $0 and this whole comparison inverts. Keep that inversion in mind; it returns in the next section.
The 5-person team: where the crossover bites. Ten concurrent default sandboxes, 8 hours a day, 22 workdays: 10 × 8 × 22 = 1,760 sandbox-hours × $0.1656 = $291.46 in usage, plus the $150 Pro fee the 8-hour sessions force (Hobby caps at 1 hour), for ~$441/month. One AX52 holds this comfortably — ten 2 vCPU sandboxes against 16 threads, 40 GiB of guest RAM against 64 GiB — at $70 flat. The delta is ~$371/month, or roughly $4,450 a year, which is the salary-adjusted cost of about one on-call rotation nobody has to staff, except inverted: it is money paid instead of staffing anything, which is precisely E2B's pitch. Whether $4,450/year buys enough "no infrastructure to operate" is a judgment call, but it should be made as a number, not a vibe.
The heavy shop: meters converge toward hardware. Fifty concurrent sandboxes around the clock is 36,500 sandbox-hours: $6,044 in usage plus the $1,150 Pro++ tier that 50-way average concurrency with peaks realistically needs, ~$7,194/month. Three AX52-class boxes absorb it for ~$210. At this scale E2B's own Enterprise tier reframes the question — $3,000/month minimum with BYOC and negotiated terms — which is the vendor telling you, via the price list, that steady-state heavy always resolves to dedicated capacity. The only question is whose.
Utilization is the whole game
The crossover (~400 sandbox-hours/month for the default shape) moves with four variables, and every team controls at least two of them:
- Idle seconds bill at full price. Agents spend most wall-clock time waiting — on model responses, on tool calls, on human approval. E2B meters the wait identically to the burn. A sandbox that computes 10 minutes of each hour but lives the full hour costs 6× its CPU cost. Auto-pausing on idle (E2B supports pause/resume with filesystem and memory snapshots) is the single highest-leverage cost control on the platform, and teams that skip it pay for it directly.
- Concurrency peaks price the plan, averages price the usage. The $150 Pro fee is set by your peak concurrent count and session length, not your total hours. A team with one daily 50-sandbox spike and near-zero baseline pays the Pro++ ladder for the spike while the usage line stays small — the worst ratio on the price list, and the workload shape where an owned box (which prices peaks at $0 marginal) looks best.
- Session length forces the tier jump. Hobby's 1-hour cap is the quiet forcing function: any workflow with sessions past an hour — long evals, overnight agents, persistent dev sandboxes — moves to Pro regardless of total hours, adding $150 before the first sandbox-second. Short, ephemeral, per-task sandboxes are E2B's home turf; long-lived ones are its tax bracket.
- Sandbox shape scales the meter linearly. The default 2 vCPU / 4 GiB is a choice, not a law. Halving to 1 vCPU / 1 GiB cuts the hourly rate from $0.1656 to ~$0.067 — and moves the crossover proportionally. Right-sizing sandboxes beats renegotiating anything.
The honest summary: E2B wins when sandbox-hours are low, spiky, or uncertain — exactly the profile of a team starting with agents. Owned capacity wins when the hours are steady and forecastable — exactly the profile of a team that succeeded with agents. The bill is a lagging indicator of adoption.
The two things money doesn't fix
No GPUs at any tier. This is the gap the TODO item names and the pricing page confirms. Modal, Daytona, and Beam all sell GPU-backed sandboxes; E2B does not, as of 2026. For a self-hosted sandbox primitive tied to a broader compute fleet, this is the structural advantage: the same Hetzner-class boxes that run CPU sandboxes today sit next to GPU nodes tomorrow, and a sandbox scheduler that already speaks Firecracker doesn't care which pool the microVM lands in. E2B can't follow you there without rebuilding its fleet — which, notably, runs today as L2 guests on GCP n1-standard-8 VMs with nested virtualization, not bare metal. You can rent E2B's orchestration; you can't rent accelerators it doesn't have.
Self-hosting E2B means operating E2B. The infrastructure repo is open source (Apache-2.0), which reads like an exit hatch until you price the operations: a Nomad + Consul + Terraform stack with Postgres, Loki, Grafana, and KVM hosts requiring nested virtualization — GCP fully supported, AWS in beta, your Hetzner fleet on your own. Teams that have tried describe it accurately: the SDK is coupled to E2B's control plane, so "self-hosted E2B" is really "operate a second platform." The official alternative is Enterprise BYOC at a $3,000/month minimum. Neither is a hobby project; both exist for regulated teams whose data can't leave their perimeter, and they should be evaluated as compliance purchases, not cost optimizations.
Which should you pick?
- Prototyping agents, spiky evals, fewer than ~400 sandbox-hours/month: E2B Hobby, then Pro. The meter is cheap, the 150 ms starts make per-task sandboxes free architecturally, and the $100 credit funds the experiment. Set idle auto-pause on day one.
- Steady team volume past the crossover, on hardware you already rent: run Firecracker on your own boxes. The marginal cost is near zero, the isolation primitive is identical, and the ~$4,500/year delta funds the small amount of automation (snapshot templates, warm pools, idle reaping) that reproduces E2B's product surface.
- GPU-adjacent agent work, or data that can't leave your perimeter: neither Hobby math nor Pro math applies. Compare Modal/Daytona GPU sandbox rates against owned GPU nodes, or price E2B Enterprise BYOC as compliance spend — and don't let CPU-sandbox pricing anchor a GPU decision.
The deeper pattern is worth naming: per-second billing plus sub-second starts made the sandbox disposable, and disposable sandboxes made agent architectures possible. That was E2B's real contribution — bigger than any rate on the price list. But disposability is a property of Firecracker, not of E2B's invoice. Once your sandbox-hours are steady enough to forecast, the same disposability runs on machines you own, next to GPUs E2B can't sell you, at a tenth of the meter. Rent the primitive to learn; own the fleet to scale.
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.
