AWS Nitro Sandboxes, benchmarked against 15 sandbox providers
Contents
AWS now lets you run hardware-isolated microVM sandboxes inside your own AWS account. We wanted to know how they compare with the sandbox providers agents already run on, so we put AWS Nitro Sandboxes through Starsling's open sandbox benchmark against 15 of them.
The short version: on AMD instances the hardware is fast, 3rd on a full Better-Auth pipeline and 4th on CPU, at about $0.24 an hour for a 4 vCPU / 8 GB sandbox. But AWS throttles every sandbox by default, and with those limits on, the same pipeline drops to 11th and a cold pnpm install takes 106 seconds instead of 12. If you're evaluating it, raise the limits before you judge it.
The full pipeline
The Better-Auth suite runs the same steps as the project's own CI: clone, install, build, lint and typecheck. That's close to what a coding agent does on every task. With AWS's per-sandbox limits switched off, Nitro finished in 148 seconds, behind only Brezel and boat. With the defaults it took 247 seconds, between tama and Vercel Sandbox.
tama ran on a much larger machine than 4 vCPU / 8 GB in Starsling's run, so its results here aren't like for like.
Why the defaults hurt
Each sandbox gets about 100 Mbps of network, 1,500 disk IOPS and 4,096 open connections by default. That's plenty for a light function, but not for a dev box. A pnpm install is thousands of small writes and hundreds of connections at once, so it hits all three limits together, and the cold install goes from 4th place to last of 17.
Disk shows the cap most clearly. Random 4K writes ran at 1,400 MB/s with the limits off, second only to Brezel, and at 6.1 MB/s with the 1,500 IOPS default. You can change each limit per pod, and you should for any agent that installs packages or builds code.
The instance type matters too
We ran the same sandbox on Intel (m8id) first. With the limits off it placed 7th on CPU and 9th on Better-Auth. On AMD (m8a) each vCPU is a full core instead of a hyperthread, and it moved up to 4th on CPU and 3rd on the pipeline.
Price
Priced as one eighth of an on-demand m8a.8xlarge, a 4 vCPU / 8 GB sandbox costs about $0.24 an hour when fully busy. That sits with Namespace and Novita, well below the $0.33 most hosted providers charge, though above boat, run.cloud and tama. It doesn't include the EKS cluster, the warm pool or the time it takes to run them yourself.
Where it fits
Nitro Sandboxes give you hardware-isolated microVMs inside your own AWS account, VPC and IAM. For an enterprise that can't send code or data to another provider, that's a big unlock.
It's also early. There's no snapshot, restore or fork yet (pausing deletes the pod and keeps its volume), cold starts take 3–4 seconds unless you build and run a warm pool yourself, and AWS says it isn't built for untrusted pod specs.
If you'd rather not run it yourself
We run sandboxes for our own company agent, so we've spent a lot of time on the parts Nitro leaves to you. Adapt Sandboxes are hardware-isolated microVMs, each an isolated computer for one agent run, up to 32 vCPU and 32 GiB. They wake in under a second and are metered per second while they run. A parked sandbox costs nothing to run, and its snapshot is there for the next run. You can use them with the Adapt company agent or call them from your own.
We wrote about how we got there, from gVisor on GKE to Firecracker, in How we run sandboxes for agents at scale. If you want to try them, get access.
How we ran it
We used Starsling's suites at 4 vCPU / 8 GB, one sandbox at a time, with 3 trials each. The other providers' results are from Starsling's 3 October 2026 run, and their prices are the published rates we checked in August and September 2026. The AWS results are our own run on the same harness, on EKS with m8a.8xlarge nodes (AMD EPYC Turin), from 9 to 11 November 2026. "Defaults" keeps AWS's per-sandbox limits. "Limits off" disables the network bandwidth, packet rate, DNS and disk throttles, and raises the connection limit to its 65,536 maximum, since AWS doesn't allow turning that one off. Starsling's data flags one provider whose sandboxes didn't match the requested size: tama ran on a host with 96 vCPU, so compare its numbers with care.
The company agent goes open source on January 12.
Join the waitlistAbout the author

Casey Manos
Brand and design @ Adapt