Is your organization AI ready? Take the 3-minute test

CreateOS Sandbox vs Microsoft Foundry

Foundry is the fit for enterprises on Azure building agents with Foundry models that want governed code execution inside Azure identity, VNet and data-residency controls. CreateOS Sandbox is the fit for agents that need a general Linux machine with any language, controlled network access and sizes past 2 vCPU / 4 GiB. Foundry's Code Interpreter is Python only, has no network and stops after an hour.

Microsoft now calls the product Microsoft Foundry Agent Service, formerly Azure AI Foundry; Agents (classic) is deprecated and retires on 31 March 2027.

Pricing at a Glance

CreateOS rates verified 17 October 2026. Microsoft Foundry pricing as published on 17 October 2026. Rates billed per GB rather than per GiB, or as bundled sizes, are quoted as published rather than converted.

CreateOS Sandbox

Lower compute rate
Compute
$0.0362 /vCPU-hr
Memory
$0.0116 /GiB-hr
No egress fees, 50 GiB bandwidth budgetvCPU stops billing while paused

Microsoft Foundry

Compute
$0.0994 /vCPU-hr
Memory
$0.0118 /GiB-hr

Microsoft's East US retail prices list Code Interpreter at $0.03 per session (global) and hosted agents at $0.0994 per vCPU-hour and $0.0118 per GiB-hour, with model tokens billed separately, as of 17 October 2026.

Where CreateOS Sandbox Fits Better

What CreateOS Sandbox brings to this matchup, with a plain note on where Microsoft Foundry has the edge.

Lower compute rates than Foundry hosted agents: $0.03616363 per vCPU-hour and $0.01159025 per GiB-hour, against $0.0994 and $0.0118.

Larger sandboxes, up to 8 vCPU / 16 GB on Enterprise, where hosted agents top out at 2 vCPU / 4 GiB.

Any language and package in the sandbox, with images like devbox:1, while Code Interpreter is Python with a fixed package set.

Pause and fork with memory state; Foundry publishes no fork, and Code Interpreter sessions end after 1 hour active.

Where Microsoft Foundry is strong

Agent sandboxing tied into Azure identity, with a Microsoft Entra ID per agent, VNet isolation, regional data residency and the Foundry model catalog, plus a no-network, Hyper-V-isolated Python sandbox enabled with a single tool flag.

Which to Pick

Pick CreateOS Sandbox if

  • Your agent needs Node.js, Go, Rust or shell tools, not only Python.
  • Your agent needs more than 2 vCPU / 4 GiB.
  • You want sandboxes that can be paused and forked rather than timed out.

Pick Microsoft Foundry if

  • You build agents on Azure with Foundry models and want code execution inside Azure boundaries.
  • You want Python analysis and charts with no network access at all.
  • You want hosted agents authenticated with Microsoft Entra ID in one of 31 Azure regions.

Head-to-Head Comparison

Every dimension side by side, checked on 17 October 2026 against each vendor's own docs and pricing pages. "Not published" means we could not find it on the vendor's pages, not that the feature is missing.

DimensionCreateOS SandboxMicrosoft Foundry
IsolationFirecracker micro-VM with its own guest kernelCode Interpreter: Hyper-V per session on Container Apps dynamic sessions. Hosted agents: per-session VM-isolated sandbox
Create and destroy by APIYes: REST API, CLI, MCP server and SDKs for TypeScript, Go, Python, Rust, C# and Java (all but TypeScript pre-1.0)Yes: Python, C#, JavaScript/TypeScript and Java SDKs, REST, Azure CLI, azd
Memory snapshot (pause and resume)Yes: disk, memory and CPU state saved; vCPU stops billing, memory and storage still billHosted agents restore $HOME and /files on resume, compute deprovisioned when idle; memory snapshot not published
ForkYes: fork a paused sandbox into many independent sandboxesNot published
Outbound network controlsOpen with no rules; deny-by-default once any host, IP, CIDR or wildcard rule exists, enforced on the host outside the VMCode Interpreter: no outbound network. Hosted agents: default not stated; customer VNet for outbound
Billing granularityPer second while running: $0.03616363/vCPU-hr, $0.01159025/GiB-hr, storage $0.0001584/GB-hrCode Interpreter per session; hosted agents per vCPU-hour and GiB-hour
Largest size8 vCPU / 16 GB on Enterprise; 8 vCPU / 8 GB on ProHosted agents 2 vCPU / 4 GiB per session; Code Interpreter not published
GPUNone; no GPU shapesNot published

Common Questions

Can Microsoft Foundry's Code Interpreter install packages or reach the internet?

No on both counts, per Microsoft's docs: Code Interpreter is Python with a fixed package set, and its dynamic sessions cannot make outbound requests. A CreateOS sandbox lets you install what you need and keeps outbound open until you add allow rules.

Do Foundry hosted agents keep files between turns?

Yes. Hosted agents keep $HOME and /files across turns and idle periods and restore them on resume, with compute deprovisioned when idle, and session data is deleted after 30 days of inactivity. CreateOS pause keeps disk, memory and CPU state, and the sandbox has no lifetime cap.

Is user data shared between sessions in Foundry Code Interpreter?

It can be. Microsoft warns that Code Interpreter used through a toolbox in a hosted agent has no user isolation, so all users in a project share the container context. On CreateOS each sandbox is its own Firecracker micro-VM, and sandboxes cannot reach each other unless placed on the same private network.

See CreateOS Sandbox for yourself

Book a demo, or explore the Sandbox and its pricing.