Dora NodaCloudflare took Containers and its Sandbox SDK to general availability on April 13, 2026, and the pricing line that matters is exactly one: CPU is billed at $0.000020 per vCPU-second of active time — $0.072 per active vCPU-hour — while an idle container costs memory and disk only. Against the E2B/Daytona baseline of $0.0504 per wall-clock vCPU-hour, that is a 7x cheaper CPU line for work that barely burns cycles, and a worse one for work that never stops. I repriced a 30,000-session agent month line by line: at a typical 10%-active bursty profile the Cloudflare model lands around $1,356 a month versus $2,484 on E2B/Daytona, roughly 45% cheaper — but the all-in crossover sits near 60% CPU-active, past which the old per-second model wins back.
The rest of this post is the receipt, the five fine-print lines that move the total, and the part no recompute can fix: it only runs on Cloudflare's fleet.
One correction up front, because the premise I started from was wrong. The TODO line that assigned this post described the new entrant as "a proprietary layer on Cloudflare's own Kubernetes." I went looking for the Kubernetes underneath and could not find it in anything Cloudflare has published: Containers are OCI images orchestrated through Cloudflare's own Durable-Objects-based control plane on Cloudflare's network, with each container paired to its own Durable Object. Proprietary layer, yes. Kubernetes, unverified. A post about concrete numbers should start with correct ones, so everything below says "proprietary orchestration on Cloudflare's fleet" and nothing else.
What actually shipped
Containers entered public beta in June 2025 and went GA on April 13, 2026, alongside the Sandbox SDK — the container-class sandbox offering that sits next to (not inside) the older V8-isolate sandbox story — as part of an "Agents Week" push aimed squarely at agent code-execution workloads. The GA terms that matter for cost math:
- Gate: Workers Paid plan, $5/month. There is no free tier for Containers; the $5 is a floor every account pays before the first session runs.
- Included monthly usage: 375 vCPU-minutes (6.25 vCPU-hours), 25 GiB-hours of memory, 200 GB-hours of disk. At fleet scale this covers roughly the first dozen sessions of the month — a rounding error, but it means small experiments start nearly free.
- Billing granularity: 10ms increments. Charges start when a request hits the container or it is started manually, and stop when the instance goes to sleep — sleep can happen automatically after a timeout, which is what makes scale-to-zero real rather than marketing.
- Instance tiers are fixed: lite (1/16 vCPU, 256 MiB, 2 GB) up through standard-4 (4 vCPU, 12 GiB, 20 GB). You cannot order 2 vCPUs with 4 GiB of RAM; the 2-vCPU tier (standard-3) ships with 8 GiB. Remember that — it comes back to bite in the worked example.
- CPU is active-only; memory and disk are provisioned. While your container is awake you pay for all of its tier's RAM whether your code touches it or not. The sleep timeout is therefore the single most important cost control on the platform.
The rate card, normalized to one unit
Cloudflare's official Containers pricing (last updated August 28, 2026) converts to hourly rates as follows, against the two baselines this topic has tracked:
| Line | Cloudflare Containers | E2B / Daytona | Modal |
|---|---|---|---|
| CPU | $0.072 / active vCPU-hr | $0.0504 / wall vCPU-hr | ~$0.14 / provisioned core-hr |
| Memory | $0.009 / provisioned GiB-hr | $0.0162 / GiB-hr | bundled per plan |
| Disk | $0.000252 / provisioned GB-hr | E2B: 10–20 GB free; Daytona: $0.000108/GiB-hr (5 GB free) | bundled per plan |
| Billing clock | per 10ms, stops at sleep | per second, while session runs | per second, provisioned |
| Floor | $5/mo Workers Paid | free tiers / credits | varies |
Two ratios fall out immediately. On CPU, Cloudflare is cheaper whenever the sandbox's CPU-active fraction is below $0.0504 / $0.072 = 70%. On memory, Cloudflare's per-unit rate is ~44% cheaper — but it accrues against the provisioned tier for as long as the container stays awake, so tier granularity and sleep behavior decide whether that discount survives contact with a real workload. Note the direction of the industry while we are here: per-second billing is now the norm across sandbox vendors, and Vercel's Sandbox has its own active-CPU line at $0.128/vCPU-hr — Cloudflare is the cheapest active-CPU price on the board, not the only one.
The worked example: 30,000 bursty sessions
The fleet: 1,000 agent code-execution sessions a day, each holding a sandbox for 30 wall-clock minutes at 2 vCPUs and 4 GiB, burning CPU 10% of the time — the classic agent shape where the model is thinking (billed to your token meter, not your sandbox) while the sandbox idles. Thirty days, 30,000 sessions.
E2B/Daytona, per session: CPU 2 × 0.5 hr × $0.0504 = $0.0504. RAM 4 × 0.5 × $0.0162 = $0.0324. Disk inside the free allowance. Total $0.0828/session → $82.80/day → $2,484/month.
Cloudflare (standard-3: 2 vCPU, 8 GiB, 16 GB), per session: CPU 2 × 0.5 × 10% × $0.072 = $0.0072. Memory 8 × 0.5 × $0.009 = $0.036 — note the forced 8 GiB, double what the workload asked for. Disk 16 × 0.5 × $0.000252 ≈ $0.002. Total $0.0452/session → $45.20/day → $1,356/month plus the $5 base. Roughly 45% cheaper, with the CPU line doing all the work ($0.0072 vs $0.0504) while memory actually costs more ($0.036 vs $0.0324) thanks to the tier step-up.
Modal, for orientation: ~$0.14 × 2 cores × 0.5 hr ≈ $0.15/session before memory → on the order of $4,500/month, roughly 3x the Cloudflare scenario. Modal's argument was never per-CPU cheapness — it is GPUs and scale-out ergonomics — but if your sandbox fleet is CPU-only code execution, this is what provisioned billing costs you.
Sensitivity: where the lead flips
The 10% figure is doing heavy lifting, so here is the same per-session total across active fractions. E2B/Daytona sit flat at $0.0828; Cloudflare follows $0.072·f + $0.038:
| CPU-active fraction | Cloudflare/session | vs E2B/Daytona |
|---|---|---|
| 5% | $0.0416 | ~50% cheaper |
| 10% | $0.0452 | ~45% cheaper |
| 25% | $0.0560 | ~32% cheaper |
| 50% | $0.0740 | ~11% cheaper |
| ~62% | ≈ $0.0828 | break-even |
| 100% ( pegged compile farm) | $0.1100 | ~33% more |
The honest headline number is the break-even, not the 45%: CPU-only parity sits at 70% active, but the forced-RAM overhang drags the all-in crossover down to about 62%. Below it, active-CPU billing wins by a lot; above it — sustained builds, video encode, anything pegged — you are paying a premium for a pricing innovation that no longer applies to you, on hardware with twice the RAM you ordered. Size your fleet's actual active fraction before migrating; anything with a bimodal workload (bursty agents plus a compile-farm tail) wants the two split across billing models, not averaged into one.
The five fine-print lines
- Sleep timeout is the cost control. Memory accrues at the full provisioned tier until the container sleeps. A container kept warm between requests for latency reasons bills $0.009/GiB-hr for the privilege — 8 GiB warm 24/7 is $51.84/month before running a line of code. Tune the idle timeout first, benchmark cold starts second.
- Tiers force over-provisioned RAM. There is no 2-vCPU/4-GiB shape; standard-3's 8 GiB is what moved the break-even from 70% to 62%. If your sandbox image is small, the lite and basic fractional tiers are the real bargain on this price list.
- Three surcharge lines ride along. Per the Sandbox pricing page, every sandbox also bills Workers (inbound requests), its own Durable Object (duration plus requests), and optionally Workers Logs. Each is small per session; none is zero across 30,000 of them. Budget ~5–10% on top for the trio combined at this volume.
- Egress is generous, not free. $0.025/GB in North America and Europe with 1 TB included monthly. Sandbox traffic is mostly ingress (pulling images and repos is free) — but agents that push large artifacts out on every run can feel this line. E2B and Daytona both include egress too, so call it a wash unless your agents exfiltrate gigabytes.
- The $5 floor inverts small scale. Fifty sessions a month cost about $2.26 in compute plus $5 base = $7.26 on Cloudflare, versus ~$4.14 on E2B before free credits. Under a few hundred sessions a month, the competitor's free tier beats the better rate card. Active-CPU billing is a fleet discount, not a hobbyist one.
The part you cannot reprice
Here is what the line-by-line math cannot change: Cloudflare Containers run only on Cloudflare's fleet, under a proprietary orchestration plane, with per-VM-isolated Linux containers that still share the host kernel — a weaker boundary than E2B's Firecracker microVMs. There is no self-host path, no bring-your-own-machine story, and no GPU tier on the price list (E2B and Daytona do not offer GPUs either; Modal owns that axis). E2B's infrastructure is open source with a documented Firecracker-on-your-cloud path, and Daytona offers self-hosting — both let an operator move the same primitive onto machines they own when the bill or the compliance posture demands it.
That is the actual fork in the road this GA forces. If your sandbox fleet is bursty, CPU-idle, and allowed to live on someone else's computers, Cloudflare just posted the best price in the category and you should take it — run the sensitivity table against your own active fraction first. If your roadmap ends with sandboxes on hardware you control — cost floors you set, kernels you chose, data that never leaves your rack — then the cheapest proprietary runtime is still a rental, and the primitive worth investing in is the one you can relocate: Firecracker microVMs on your own node pool, billed to yourself at the power bill.
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.
