Fly Sprites vs Northflank: pricing, isolation and limits

Sprites bill measured CPU and memory in use; Northflank bills allocated vCPUs and memory at lower rates, so it costs less here.

Metering CPU is not enough on its own: Runtime, which meters CPU and charges little for memory, costs 81% less than Sprites and 42% less than Northflank for a waiting agent. 1,000 one-minute runs of a 2 vCPU, 4 GiB sandbox cost $3.31 on Fly Sprites, $1.11 on Northflank and $0.64 on Runtime. Runtime's CPU rate is about a third of Sprites', and its memory rate is below both. Rates checked 23 September 2026.

At a glance

Northflank's isolation and regions are from its sandbox page, read 25 September 2026.

	Fly Sprites	Northflank	Runtime
Isolation	Firecracker microVM	A microVM with Kata Containers, or gVisor	Firecracker microVM, own kernel
CPU	$0.07 per CPU-hour of measured CPU	$0.01667 per allocated vCPU-hour	$0.025 per vCPU-hour of measured CPU, with a small floor
Memory	$0.04375 per GB-hour of memory in use	$0.00833 per GB-hour	$0.0075 per reserved GiB-hour
Idle	Pauses itself about 30 seconds after the last activity	Ephemeral by default; volumes and databases for state	Pause keeps files, memory and processes, 1 to 365 days
Lifetime	Persistent; the disk survives every pause	No forced time limit, in Northflank's words	Leases extended as often as needed, or persistent: true
Plan fee	None required; optional monthly plans	None published for compute	None; prepaid credit from $10
Free start	$30 of trial credit	2 free services and 1 free database	50 sandbox hours, no card
Sandbox model	A Sprite, created by name	A service with a deployment plan	A sandbox, created in one call

Is paying only for the CPU you use always cheaper?

No. What matters is the price of each meter, and memory is where these two part company. A Sprite's memory costs $0.04375 per GB-hour; Northflank's costs $0.00833, about a fifth. In the example below, a Sprite's 4 GB of memory alone costs $2.92 per 1,000 runs, where Northflank's whole bill, CPU and memory, is $1.11.

Metering CPU helps Sprites only on the CPU line: $0.39 for 20 CPU-seconds a run against Northflank's $0.556 for two allocated vCPUs. With both CPUs busy the whole minute, Sprites' CPU line rises to $2.33 and Northflank's stays where it was.

Runtime pairs a metered CPU at $0.025 a vCPU-hour with the lowest memory rate of the three, $0.0075 per GiB-hour. That makes it the cheapest while an agent waits. Northflank's allocated vCPU rate is lower still, so for work that keeps every CPU busy the whole time, Northflank costs less.

Cost for the same job

Priced for 1,000 runs of 60 seconds on 2 vCPU and 4 GiB (4 GB on each rival), with 20 CPU-seconds of computing per run. The Sprite is billed as though all 4 GB are in use for the full minute.

TextSprites     CPU    1,000 × 20 s / 3,600 × $0.07          = $0.39            Memory 1,000 × 60 s / 3,600 × 4 × $0.04375   = $2.92            Total                                           $3.31Northflank  CPU    1,000 × 60 s / 3,600 × 2 × $0.01667   = $0.556            Memory 1,000 × 60 s / 3,600 × 4 × $0.00833   = $0.555            Total                                           $1.11Runtime     CPU    1,000 × 20 s / 3,600 × $0.025         = $0.14            Memory 1,000 × 60 s / 3,600 × 4 × $0.0075    = $0.50            Total                                           $0.64

A month of 100,000 runs adds up to $330.56 on Sprites, $111.10 on Northflank and $63.89 on Runtime. With both CPUs busy for the whole minute, it is $1.11 on Northflank, $1.33 on Runtime and $5.25 on Sprites, so for code that never pauses Northflank's allocated rate is the cheapest. A Sprite using just 1 GB would pay $1.12 for the waiting job. Plan fees, storage, network and free allowances are left out.

Which one to pick

• Fly Sprites, if each agent should own a lasting machine: a 100 GB disk that is cheap to keep while the Sprite sleeps, and checkpoints of the filesystem that take about a second.
• Northflank, for CPU-bound jobs, GPUs as large as H100, or sandboxes placed in your own AWS, GCP, Azure or Oracle account.
• Runtime, for agents that wait on a model: the lowest price of the three for that pattern, a pause that keeps memory and processes, and forks of a running sandbox.

Moving to Runtime

A Sprite you address by name maps to a named Runtime sandbox, which is woken if paused and created if it does not exist yet:

Pythonfrom withruntime import Sandboxdev = Sandbox.get_or_create("agent-42", idle_pause_seconds=300)dev.exec("git pull || true")

Other Sprites calls port directly: client.createSprite(name) is Sandbox.create() and sprite.execFile(cmd, args) is exec([cmd, ...args]). On Northflank, the service and deployment plan disappear into one create call. npx withruntime compare --from fly or --from northflank reprices past usage; npx withruntime switch --from fly before the first top-up adds matching credit, up to $100. A coding agent can make the change from the one instruction in migration, on a branch and on the free trial. See Runtime vs Fly.io and Runtime vs Northflank.

Sources

Rates checked 23 September 2026; the Sprites, Sprites lifecycle and Northflank sandbox pages read 25 September 2026.

• Fly.io pricing, Sprites, Sprites lifecycle
• Northflank pricing, Northflank sandboxes, Sandboxes on Northflank
• Runtime pricing and Python SDK

Facts on this page were checked on 25 September 2026.