Northflank vs Runloop Northflank is 79% cheaper per agent run

Which is better, Northflank or Runloop?

It depends on whether you need raw compute or a benchmarking service. For agents that spend their time on model calls, Runtime bills measured CPU, so the 1,000-run job costs $0.64 there, against $1.11 on Northflank and $5.33 on Runloop. Pick Northflank for sandboxes that keep every CPU busy, where it is cheapest, for GPUs billed by the second, or for your own cloud account. Pick Runloop for evaluation and RL rollouts on your own workload, with stated SOC 2, HIPAA and GDPR readiness.

At a glance

Which is cheaper for a sandbox that runs all day?

Northflank, by a wide margin. Both bill every allocated CPU and GB for as long as the sandbox runs. Runloop's CPU rate is about six and a half times Northflank's ($0.108 against $0.01667), and its memory rate is about three times ($0.0252 against $0.00833). For one 2 vCPU, 4 GB sandbox left running for a 30-day month of 720 hours:

Runloop's figure leaves out the $250-a-month Pro plan, which suspend and resume need. An agent sandbox spends most of its hours waiting, and there Runtime is the cheapest: it bills the floor of a twentieth of a vCPU rather than two whole vCPUs.

Cost for the same job

Short runs instead of an all-day box: 1,000 of them, 60 seconds each, with 20 CPU-seconds of work, on 2 vCPU and 4 GiB. Northflank is sized at 4 GB. Runloop's MEDIUM devbox has 2 CPUs, 4 GB and 8 GB of disk.

At 100,000 runs a month, the totals are $111.10 on Northflank, $532.56 on Runloop and $63.89 on Runtime. With both CPUs busy for the whole minute, the order at the top changes: $1.11 on Northflank, $1.33 on Runtime, $5.33 on Runloop. When every CPU works all the time, Northflank's low allocated rate is the cheapest way to pay. Plan fees, idle storage, network, taxes and free allowances are not included.

Which one to pick

• Northflank for long-running sandboxes that keep every CPU busy, for L4, A100 and H100 GPUs billed by the second, or for sandboxes placed in your own cloud account alongside your other services.
• Runloop for an evaluation platform: models evaluated on your own workload, RL rollouts, and stated SOC 2, HIPAA and GDPR readiness.
• Runtime for agents that spend their time on model calls: 42% below Northflank and 88% below Runloop on the typical run, a Firecracker microVM per sandbox, pause and forks that keep memory, and guardrails such as a daily spending limit per key. Benchmarks run here too: agent evals and SWE-bench.

Moving a long-lived sandbox

There is no direct import for either; follow map the calls. A Northflank service kept up for weeks, or a Runloop devbox that never goes away, maps to a persistent Runtime sandbox with a cap on its lifetime spend:

TypeScriptimport { Sandbox } from "withruntime";// Runs while credit lasts; restart() brings it back after a stop.const server = await Sandbox.create({  persistent: true,  maxTotalCostMicros: 50_000_000, // at most $50 over its life});await server.update({ persistent: false }); // back to an ordinary sandboxawait server.stop();

Pythonfrom withruntime import Sandboxserver = Sandbox.create(persistent=True, max_total_cost_micros=50_000_000)server.update(persistent=False)  # back to an ordinary sandboxserver.stop()

To compare on your own numbers, run npx withruntime compare --from northflank or --from runloop. It reprices what you ran. The one instruction in migration lets a coding agent do the port on a branch and test it on the free usage. See also Runtime vs Northflank and Runtime vs Runloop.

Sources

Rates checked 25 October 2026. Northflank's pricing and sandbox pages and Runloop's pricing page and devbox sizes reread 26 October 2026.

• Northflank pricing, Northflank sandboxes, Sandboxes on Northflank
• Runloop pricing, Devbox sizes, Devbox lifecycle
• Runtime pricing and JavaScript SDK

Facts on this page were checked on 26 October 2026.