E2B
E2B (FoundryLabs, Inc.)
- Isolation
- Firecracker microVM per sandbox
- Max session
- 1 hour (Hobby), 24 hours (Pro), custom (Enterprise) of continuous runtime; pausing resets the window
- Status
- GAinferred
Agent code sandboxes - verified
AI agent code sandboxes compared
We read the docs and pricing pages of 18 code-execution sandboxes for AI agents on October 18, 2026, plus 2 open-source runtimes you can host yourself, and kept 280 verbatim quotes. They come in three kinds: 11 hosted sandbox platforms, 4 cloud primitives and 3 model-provider built-in tools. Prices stay in each provider's own unit, so they are grouped by unit and never ranked across units.
- 3 products state a Firecracker microVM per sandbox: E2B, Vercel Sandbox and CodeSandbox SDK.
- 3 run with no outbound network by default: Anthropic code execution tool (Claude API), Google Gemini Enterprise Agent Platform Code Execution (formerly Vertex AI Agent Engine) and Azure Container Apps dynamic sessions.
- GPU is stated by 4 products: Daytona, Modal Sandboxes, Runloop Devboxes and Northflank Sandboxes.
Want a verdict instead of a table? Read E2B vs Daytona vs Modal or how to choose a runtime boundary.
- Hosted platforms
- 11
- Cloud primitives
- 4
- Built-in tools
- 3
- Verified quotes
- 280
Three kinds of sandbox
A hosted platform, a primitive inside a cloud you already use, or a tool your model provider runs for you. They solve the same problem with different trade-offs, so each kind is listed on its own.
Hosted sandbox platforms
11Products whose job is running agent code: you call an SDK or API, get an isolated machine, run commands, and pay for the resources it uses.
Daytona
Daytona Platforms Inc.
- Isolation
- Container class: isolated container with dedicated namespaces
- Max session
- No fixed cap stated; auto-stop can be disabled so a sandbox runs indefinitely
- Status
- GAinferred
Modal Sandboxes
Modal Labs
- Isolation
- gVisor runtime
- Max session
- Default 5 minutes, configurable up to 24 hours
- Status
- GAinferred
Vercel Sandbox
Vercel
- Isolation
- Firecracker microVM per sandbox
- Max session
- 45 minutes (Hobby), 24 hours (Pro, Enterprise) per session; persistent sandboxes resume in new sessions
- Status
- GAinferred
Runloop Devboxes
Runloop AI, Inc.
- Isolation
- Isolated ephemeral virtual machines
- Max session
- Default max lifetime 1 hour, configurable at creation
- Status
- GAinferred
Northflank Sandboxes
Northflank
- Isolation
- microVM per workload with Kata Containers or gVisor
- Max session
- No forced time limits stated
- Status
- GAinferred
CodeSandbox SDK
CodeSandbox (part of Together AI)
- Isolation
- Firecracker-powered virtual machines
- Max session
- Unlimited session length on all plans
- Status
- GAinferred
Blaxel Sandboxes
Blaxel
- Isolation
- microVMs (Mark 3.0 infrastructure)
- Max session
- Not stated
- Status
- Sandboxes GA; next-gen Carbon infrastructure in private previewinferred
Morph Cloud
Morph Labs
- Isolation
- VMs (instances are running VMs started from snapshots)
- Max session
- Not stated
- Status
- Not stated
Beam Sandboxes
Beam
- Isolation
- Container
- Max session
- Default 600-second lifetime; keep_warm_seconds=-1 for manual termination
- Status
- GAinferred
Fly.io Sprites
Fly.io
- Isolation
- Hardware-isolated
- Max session
- Not stated
- Status
- GAinferred
Cloud primitives
4Sandbox or code-execution building blocks inside a larger cloud, billed and configured the way the rest of that cloud is.
- Isolation
- Containers: microVM per instance with own kernel and network
- Max session
- Not stated
- Status
- Public beta for the sandbox container scheduling policy
Amazon Bedrock AgentCore Code Interpreter
Amazon Web Services
- Isolation
- Containerized environment
- Max session
- Default 15 minutes, extendable up to 8 hours
- Status
- GAinferred
Google Gemini Enterprise Agent Platform Code Execution (formerly Vertex AI Agent Engine)
Google Cloud
- Isolation
- Not stated
- Max session
- Not stated
- Status
- Not stated
- Isolation
- Hyper-V isolation
- Max session
- Each code execution limited to 220 seconds
- Status
- GAinferred
Model-provider built-in tools
3Code execution the model calls itself through the provider's API. You do not manage the machine; the model writes and runs the code.
- Isolation
- Fully sandboxed virtual machine
- Max session
- Not stated
- Status
- GAinferred
- Isolation
- Linux-based container, isolated from host and other containers
- Max session
- Containers expire 30 days after creation
- Status
- GAinferred
- Isolation
- Not stated
- Max session
- 30 seconds maximum runtime
- Status
- Not stated
Feature matrix
Every value links to the page it came from. Not stated means the checked pages did not give it; hover for the recorded note. Scroll the tables sideways on a small screen.
Runtime and lifecycle
Capabilities and deployment
Pricing by unit
Each block holds one unit, as the provider states it. Rows are sorted by price inside a block only; a price in one block is never compared with a price in another, and nothing is converted.
CPU
Five different CPU units. A physical core is stated as 2 vCPU by Modal and Beam, an Active CPU hour excludes time waiting on I/O, and a CPU-hour is the provider's own wording. Each unit is its own block and none is converted into another.
| Sandbox | Price | Unit, as stated | Quote and source | Notes |
|---|---|---|---|---|
| Per physical core-second | ||||
| Beam Sandboxes | $0.0000375 | per second per physical core (2 vCPU equivalent), Sandboxes | “$0.0000375/sec · per core” | - |
| Modal Sandboxes | $0.00003942 | per physical core per second (Sandbox rate; 1 physical core = 2 vCPU equivalent) | “0.00003942/ core / sec” | Minimum 0.125 cores per container. |
| Per vCPU-second | ||||
| E2B | $0.000014 | per second, 1 vCPU (Hobby / Pro) | “1 · Hobby / Pro · $0.000014/s” | Table lists 1, 2 (default), 4, 6, 8 vCPU rows scaling linearly; 2 vCPU is $0.000028/s. |
| Cloudflare Sandboxes (Containers and Dynamic Workers) | $0.000020 | per additional vCPU-second (after 375 vCPU-minutes/month) | “375 vCPU-minutes/month + $0.000020 per additional vCPU-second” | CPU billed on active usage. |
| Per Active CPU hour | ||||
| Vercel Sandbox | $0.128 | per Active CPU hour (Pro/Enterprise, iad1) | “Sandbox Active CPU · 5 hours/month · $0.128/hour · $0.128/hour” | Hobby includes 5 hours/month. Rates vary by region; table rates are for iad1. |
| Per vCPU-hour | ||||
| Northflank Sandboxes | $0.01667 | per vCPU-hour | “$0.01667/ vCPU / hour” | Platform-wide Pricing page is the general Northflank compute price list; the sandboxes page repeats the $0.01667/vCPU-hour CPU rate. |
| Daytona | $0.0504 | per vCPU-hour | “vCPU $0.0504/h” | - |
| Daytona | $0.0858 | per vCPU-hour, Windows OS | “Windows $0.0858/vCPU/h” | - |
| Amazon Bedrock AgentCore Code Interpreter | $0.0895 | per vCPU-hour (active consumption) | “Code Interpreter · CPU · $0.0895 per vCPU-hour; Memory · $0.00945 per GB-hour” | - |
| Per CPU-hour | ||||
| Fly.io Sprites | $0.0385 | per CPU-hour (cumulative CPU usage) | “$0.0385 /CPU-hour” | - |
| Runloop Devboxes | $0.108 | per CPU-hour (devbox) | “Devbox CPU $0.108 / CPU-hr” | - |
Memory
GiB and GB are different units, and per second is not per hour, so each stays in its own block. Blaxel's memory rate includes CPU.
| Sandbox | Price | Unit, as stated | Quote and source | Notes |
|---|---|---|---|---|
| Per GB RAM-second (CPU included) | ||||
| Blaxel Sandboxes | $0.0000115 | per GB RAM-second of active time (CPU included, scales with memory) | “$0.0000115 / GB RAM · second” | - |
| Per GiB-second | ||||
| Cloudflare Sandboxes (Containers and Dynamic Workers) | $0.0000025 | per additional GiB-second memory (after 25 GiB-hours/month) | “25 GiB-hours/month included +$0.0000025 per additional GiB-second” | Memory billed on provisioned instance size. |
| E2B | $0.0000045 | per second, 1 GiB memory (Hobby / Pro) | “1 · Hobby / Pro · $0.0000045/s” | Memory rows 1, 2, 4 (default), 8 GiB scale linearly. |
| Beam Sandboxes | $0.0000064 | per second per GiB RAM, Sandboxes | “$0.0000064/sec · per GiB” | - |
| Modal Sandboxes | $0.00000667 | per GiB memory per second (Sandbox rate) | “0.00000667/ GiB / sec” | - |
| Per GiB-hour | ||||
| Daytona | $0.0162 | per GiB memory-hour | “Memory GiB $0.0162/h” | - |
| Per GB-hour | ||||
| Northflank Sandboxes | $0.00833 | per GB-hour memory | “$0.00833/ GB / hour” | Platform-wide Pricing page is the general Northflank compute price list; the sandboxes page repeats the $0.01667/vCPU-hour CPU rate. |
| Amazon Bedrock AgentCore Code Interpreter | $0.00945 | per GB-hour memory (active consumption) | “Code Interpreter · CPU · $0.0895 per vCPU-hour; Memory · $0.00945 per GB-hour” | - |
| Vercel Sandbox | $0.0212 | per provisioned memory GB-hour (Pro/Enterprise, iad1) | “Sandbox Provisioned Memory · 420 GB-hours/month · $0.0212/GB-hour · $0.0212/GB-hour” | - |
| Fly.io Sprites | $0.021875 | per GB-hour memory (actual usage) | “$0.021875 /GB-hour” | - |
| Runloop Devboxes | $0.0252 | per GB-hour memory (devbox) | “Devbox Memory $0.0252 / GB-hr” | - |
Session and container time
How the model-provider tools and Azure's built-in sessions bill: per container session or per hour of a session or container, as each provider states it.
| Sandbox | Price | Unit, as stated | Quote and source | Notes |
|---|---|---|---|---|
| Per 20-minute session per container | ||||
| OpenAI Code Interpreter (Responses API containers) | $0.03 | per 20-minute session per container, 1 GB | “1 GB $0.03, 4 GB $0.12, 16 GB $0.48, 64 GB $1.92 per 20-minute session per container.” | - |
| OpenAI Code Interpreter (Responses API containers) | $0.12 | per 20-minute session per container, 4 GB | “1 GB $0.03, 4 GB $0.12, 16 GB $0.48, 64 GB $1.92 per 20-minute session per container.” | - |
| OpenAI Code Interpreter (Responses API containers) | $0.48 | per 20-minute session per container, 16 GB | “1 GB $0.03, 4 GB $0.12, 16 GB $0.48, 64 GB $1.92 per 20-minute session per container.” | - |
| OpenAI Code Interpreter (Responses API containers) | $1.92 | per 20-minute session per container, 64 GB | “1 GB $0.03, 4 GB $0.12, 16 GB $0.48, 64 GB $1.92 per 20-minute session per container.” | - |
| Per hour per container | ||||
| Anthropic code execution tool (Claude API) | $0.05 | per hour per container (beyond free hours) | “Additional usage beyond 1,550 hours is billed at $0.05 USD per hour, per container” | Execution time has a 5-minute minimum. |
| Per session-hour | ||||
| Azure Container Apps dynamic sessions | $0.03 | per session-hour, platform-managed built-in container (Python, Node.js, Shell), pay as you go | “Platform-managed built-in container (Python, Node.js, Shell) · $0.03 per session-hour” | - |
GPU
GPU rates the pricing pages list for sandboxes or the provider's general compute.
| Sandbox | Price | Unit, as stated | Quote and source | Notes |
|---|---|---|---|---|
| Per hour, one GPU | ||||
| Daytona | $2.27 | per hour, Nvidia H100, preemptible | “Nvidia H100 $2.27/h” | Other GPUs listed: B300, B200, AMD MI355X, H200, RTX PRO 6000, RTX 5090, RTX 4090. |
| Northflank Sandboxes | $2.74 | per hour, NVIDIA H100 80GB | “NVIDIA H100 80GB $2.74 / hour” | Also listed: L4, A100 40GB/80GB, RTX PRO 6000. Platform-wide Pricing page is the general Northflank compute price list; the sandboxes page repeats the $0.01667/vCPU-hour CPU rate. |
| Daytona | $3.95 | per hour, Nvidia H100, on-demand | “Nvidia H100 $3.95/h” | - |
Storage and snapshots
Disk, volume, snapshot and image storage, each in the provider's own unit.
| Sandbox | Price | Unit, as stated | Quote and source | Notes |
|---|---|---|---|---|
| Included storage | ||||
| E2B | Free | storage, 10 GiB on Hobby | “10 · Hobby · Free” | - |
| E2B | Free | storage, 20 GiB on Pro | “20 · Pro · Free” | - |
| Per GiB-hour | ||||
| Daytona | $0.000108 | per GiB storage-hour (after first 5 GiB free) | “Storage GiB $0.000108/h” | Page footnote: 'Price per GiB after first 5 free'. |
| Per GB-hour | ||||
| Fly.io Sprites | $0.000027 | per GB-hour cold storage | “COLD $0.000027 /GB-hour” | - |
| Runloop Devboxes | $0.000072 | per GB-hour snapshot storage | “Snapshot Storage $0.000072 / GB-hr” | - |
| Runloop Devboxes | $0.00034236 | per GB-hour devbox storage | “Devbox Storage $0.00034236 / GB-hr” | - |
| Fly.io Sprites | $0.000683 | per GB-hour hot storage | “HOT $0.000683 /GB-hour” | - |
| Per GB-second | ||||
| Cloudflare Sandboxes (Containers and Dynamic Workers) | $0.00000007 | per additional GB-second disk (after 200 GB-hours/month) | “200 GB-hours/month +$0.00000007 per additional GB-second” | - |
| Per GB-month | ||||
| Blaxel Sandboxes | $0.045 | per GB-month images | “$0.045 / GB · month” | - |
| Vercel Sandbox | $0.05 | per GB-month drive storage | “Drive Storage · 15 GB (lifetime) · $0.05/GB-month · $0.05/GB-month” | - |
| Vercel Sandbox | $0.08 | per GB-month snapshot storage | “Snapshot Storage · 15 GB (lifetime) · $0.08/GB-month · $0.08/GB-month” | - |
| Blaxel Sandboxes | $0.12 | per GB-month provisioned volume capacity | “$0.12 / GB · month” | - |
| Blaxel Sandboxes | $0.20 | per GB-month snapshot storage | “Memory and filesystem snapshots that keep suspended sandboxes ready to resume. · $0.20 / GB · month” | - |
Network egress
Outbound data transfer. Some of these are the provider's general network price rather than a sandbox-specific one, as marked.
| Sandbox | Price | Unit, as stated | Quote and source | Notes |
|---|---|---|---|---|
| Per GiB | ||||
| Modal Sandboxes | $0.04 | per GiB network egress | “Network Egress$0.040.04/ GiB” | Starter includes 1 TiB/month free egress. Platform-wide Listed on Modal's general pricing page; unlike its CPU and memory rows it is not labelled a Sandbox rate, so it is the platform-wide egress price. |
| Per GB | ||||
| Cloudflare Sandboxes (Containers and Dynamic Workers) | $0.025 | per GB egress, North America and Europe (1 TB/month included) | “North America & Europe · $0.025 · 1 TB” | - |
| Northflank Sandboxes | $0.06 | per GB network egress | “Network Egress $0.06/ GB” | Platform-wide Pricing page is the general Northflank compute price list; the sandboxes page repeats the $0.01667/vCPU-hour CPU rate. |
| Vercel Sandbox | $0.15 | per GB data transfer (Enterprise; Pro included in Flat Rate CDN) | “Sandbox Data Transfer · 20 GB/month · Included in Flat Rate CDN · $0.15/GB” | - |
Sandbox creations
A per-creation charge on top of compute.
| Sandbox | Price | Unit, as stated | Quote and source | Notes |
|---|---|---|---|---|
| Per 1M sandbox creations | ||||
| Vercel Sandbox | $0.60 | per 1M sandbox creations | “Sandbox Creations · 5,000/month · $0.60/1M · $0.60/1M” | - |
Credits
CodeSandbox bills VM runtime in credits. Its pricing page states two figures that do not reconcile; both are shown and neither is picked.
| Sandbox | Price | Unit, as stated | Quote and source | Notes |
|---|---|---|---|---|
| Per VM credit | ||||
| CodeSandbox SDK | $0.015 | per VM credit (credits measure VM runtime) | “They are priced at $0.015 each.” | The page text does not include the per-VM-tier credits-per-hour table. |
| Per hour, on-demand VM credits | ||||
| CodeSandbox SDK | $0.15 | per hour, on-demand VM credits on Scale (as stated; unit wording ambiguous) | “priced at $0.15 per hour” | - |
Plan fees
Monthly plan prices and minimums, usually on top of usage. A plan fee is not a sandbox rate and is not compared with one.
| Sandbox | Price | Unit, as stated | Quote and source | Notes |
|---|---|---|---|---|
| Per month | ||||
| Cloudflare Sandboxes (Containers and Dynamic Workers) | $5 | per month, Workers Paid plan (required; includes monthly usage) | “with included monthly usage as part of the $5 USD per month Workers Paid plan” | - |
| Beam Sandboxes | $20 | per month, Developer plan (includes $20 credits) | “Developer $20/month $20 in credits each month” | - |
| Fly.io Sprites | $100 | per month, Hero plan (1,200 CPU-hours, 4,800 RAM GB-hours, 150 GB storage included) | “Hero ($100/mo) includes 1,200 CPU-hours, 4,800 RAM GB-hours, and 150 GB of storage” | - |
| E2B | $150 | per month, Pro plan base price (plus usage) | “Base price · $0/month · $150/month · Custom” | - |
| CodeSandbox SDK | $170 | per month per workspace, Scale plan (from) | “From $170 per month per workspace” | - |
| Runloop Devboxes | $250 | per month, Pro plan (plus usage) | “$250 /month + usage” | - |
| E2B | $3,000 | per month minimum, Enterprise | “Enterprise requires $3,000/mo minimum.” | - |
Billed another way
Charges with no rate of their own on the checked page: billed at other rates, as model tokens, by resource or as a management fee.
| Sandbox | Price | Unit, as stated | Quote and source | Notes |
|---|---|---|---|---|
| No rate of its own | ||||
| Beam Sandboxes | $0.019 per vCPU-hour + $0.009 per GB-hour | BYOC management fee | “You pay Beam a flat management fee of $0.019/hr per vCPU and $0.009/hr per GB of RAM” | - |
| Runloop Devboxes | $0.252 | per hour blueprint build | “Blueprint Build $0.252 / hour” | - |
| Azure Container Apps dynamic sessions | Resource-based | custom container sessions billed by session pool resources | “Custom container sessions are billed based on the resources consumed by the session pool.” | - |
| Gemini API code execution | No additional charge | billed as model input/output tokens | “There's no additional charge for enabling code execution from the Gemini API.” | - |
| Modal Sandboxes | Standard GPU rates | GPU on Sandboxes billed at standard resource rates | “GPU See standard pricing” | - |
No public price
Entries whose checked pages state no price, with the reason recorded for each.
E2B Runtime / E2B Embed (open source)
Not stated - Open-source software; no price stated. Infrastructure costs are not covered by the checked pages.
Daytona (archived open-source repository)
Not stated - Open-source software; no price stated.
Morph Cloud
Not stated - Docs state usage is measured in Morph Compute Units (MCUs) and that prices are shown only in the logged-in billing console; no public per-unit price found.
“Morph Cloud measures resource usage in Morph Compute Units (MCUs). MCUs account for resource consumption; they are not a fixed number of instance hours.”
“The billing console lists the plans available to your organization, including prices, included credits, renewal terms, and resource limits.”
Google Gemini Enterprise Agent Platform Code Execution (formerly Vertex AI Agent Engine)
Not stated - Not recorded here: code execution on Gemini Enterprise Agent Platform is billed through the Vertex AI pricing page, which this dataset does not cover yet. Check that page directly.
Free tiers and credits
What each product gives away before you pay, in its own words. Not ranked: credits, quotas and free hours differ.
| Sandbox | Free tier or credits | Quote and source |
|---|---|---|
| E2B | Hobby plan: $100 one-time usage credit, no credit card required | “$100 one-time usage credit” |
| Daytona | $200 in free compute included; no credit card for trial | “$200 in free compute included.” |
| Modal Sandboxes | Starter plan: $30/month free credits, applies to Sandboxes | “$30 / month free credits” |
| Vercel Sandbox | Hobby: monthly allotment (5 Active CPU hours, 420 GB-hours memory, 5,000 creations) at no cost | “Sandbox is free for Hobby users within the usage quotas detailed above.” |
| Runloop Devboxes | Free Pro trial with $50 usage credit; Basic plan free plus usage | “$50 Usage Credit” |
| Northflank Sandboxes | Not stated | - |
| CodeSandbox SDK | Build plan $0 with SDK lite limited to 10 concurrent VM Sandboxes | “Programmatically create and manage Sandboxes at scale, but limited to 10 concurrent VM Sandboxes.” |
| Blaxel Sandboxes | Up to $200 in free credits | “Up to $200 in free credits” |
| Morph Cloud | Not statedDocs defer free credits to the billing console. | - |
| Beam Sandboxes | Developer plan includes $20 credits monthly (paid plan) | “Developer $20/month $20 in credits each month” |
| Fly.io Sprites | $30 trial credit | “$30 in trial credits” |
| Cloudflare Sandboxes (Containers and Dynamic Workers) | Not on Free plan; Sandbox requires Workers Paid | “Available on Workers Paid plan” |
| Amazon Bedrock AgentCore Code Interpreter | New AWS customers: up to $200 Free Tier credits (account-wide) | “New AWS customers receive up to $200 in Free Tier credits.” |
| Google Gemini Enterprise Agent Platform Code Execution (formerly Vertex AI Agent Engine) | Not stated | - |
| Azure Container Apps dynamic sessions | Not stated | - |
| OpenAI Code Interpreter (Responses API containers) | Not stated | - |
| Anthropic code execution tool (Claude API) | 1,550 free hours per organization per month | “Each organization receives 1,550 free hours of usage per month” |
| Gemini API code execution | Not statedNot stated on this page. | - |
Vendor-stated, not measured
Startup and resume times
What each vendor says on its own page. We did not measure these, and vendors time different things: a start from a snapshot, a resume from standby, a prewarmed pool or a filesystem checkpoint.
| Sandbox | Claim | Quote and source |
|---|---|---|
| E2B | ~150 ms startup from a snapshot or custom templateVendor-stated | “Startup · ~150 ms from a snapshot or custom template” |
| Daytona | Under 90 ms from code to executionVendor-stated | “spinning up in under 90ms from code to execution” |
| Modal Sandboxes | Containers spin up and down in 1-2 seconds (platform-wide claim, not Sandbox-specific)Vendor-stated | “Modal spins containers up and down in 1-2 seconds.” |
| Vercel Sandbox | Sandboxes start in millisecondsVendor-stated | “Sandboxes start in milliseconds, making them ideal for real-time user interactions and latency-sensitive workloads.” |
| Runloop Devboxes | Startup to first command in a few secondsVendor-stated | “Startup to running your first command takes a few seconds.” |
| Northflank Sandboxes | microVM boot in under a secondVendor-stated | “Boot a microVM in under a second.” |
| CodeSandbox SDK | Resume from memory snapshot typically 0.5-2 secondsVendor-stated | “From memory/disk snapshot: 0.5–2 seconds” |
| CodeSandbox SDK | New Sandbox from template ready in 1-3 secondsVendor-stated | “the new Sandbox is ready in 1–3 seconds” |
| Blaxel Sandboxes | Resume from standby in under 25 msVendor-stated | “Sandboxes resume from standby in under 25 milliseconds” |
| Blaxel Sandboxes | Resumes full memory state in 25 msVendor-stated | “resumes the full memory state in just 25 milliseconds” |
| Fly.io Sprites | Filesystem checkpoint in about a second (checkpoint, not cold start)Vendor-stated | “Snapshot the whole filesystem in about a second.” |
| Google Gemini Enterprise Agent Platform Code Execution (formerly Vertex AI Agent Engine) | Create and execute in under a secondVendor-stated | “Sandboxes can be created and execute code in under a second.” |
| Azure Container Apps dynamic sessions | Subsecond launch from prewarmed poolsVendor-stated | “Prewarmed pools enable subsecond launch times for interactive workloads.” |
Where the docs disagree
Two places where a provider's own pages say different things. Both sides are shown; neither is picked.
Docs disagree
Fly.io Sprites: Billed hourly or per second
“All resources are billed hourly, based on actual usage.”
Research note: Docs disagree: the pricing page says "All resources are billed hourly, based on actual usage.", while the docs comparison table (Serverless Functions vs Sprites) gives Sprites per-second billing ("Per-second"). Recorded as unresolved.
Docs disagree
CodeSandbox SDK: What a VM credit costs
“They are priced at $0.015 each.”
“priced at $0.15 per hour”
Research note: The page states $0.015 per VM credit and $0.15 per hour for on-demand VM credits on Scale (as stated; unit wording ambiguous). The page text does not include the per-VM-tier credits-per-hour table, so the two figures cannot be reconciled from it.
Self-host and open source
Runtime code you can run on your own hardware. No price is stated; you pay for the machines.
Apache-2.0 (Go)
“It is written in Go, licensed under Apache-2.0, and built by E2B.”
E2B Embed runs the whole stack on a single Linux host with KVM
“E2B Embed is the whole stack, real Firecracker sandboxes included, on a single Linux host with KVM.”
From the E2B record:
“e2b-dev/E2B is licensed under the Apache License 2.0”
Daytona (archived open-source repository)
Daytona - Unmaintained
AGPL-3.0 (GNU Affero General Public License) at tag v0.190.0
“GNU AFFERO GENERAL PUBLIC LICENSE”
Repository no longer maintained
“This repository is no longer maintained.”
Public code remains free to use and fork, as is and without support
“It remains public and free to use, fork, and build on under the LICENSE, as is and without support or warranty.”
“This repository will receive no further updates, fixes, or releases.”
From the Daytona record:
“As of June 2026, Daytona's core development has moved to a private codebase.”
Method
How to read this directory
Read from each provider's own docs, pricing pages and repositories on , with every quote string-matched against the page that day. Prices and limits change; check the linked source before you commit.
- Not a benchmark
- Nothing here measures speed, cost per task or reliability. Every value is a first-party fact from the provider's own docs or pricing page, and every startup or resume figure is labelled vendor-stated.
- Prices in the provider's own unit
- A vCPU-second, a physical core-second, an Active CPU hour, a 20-minute container session and a credit are different units. Each unit is its own block, rows are sorted only inside a block, and nothing is converted.
- Inferred means implied, not labelled
- Most products have no GA badge; an inferred status is read from published paid pricing and carries a visible inferred tag.
- Not stated means not stated
- When the checked pages do not give a value, the cell says Not stated with the note recorded for it, instead of guessing.
FAQ
What is an AI agent code sandbox?
An isolated machine where an agent can run the code it writes without touching your own systems. This directory covers 11 hosted sandbox platforms you call through an SDK, 4 cloud primitives that sit inside a larger cloud, and 3 code execution tools a model provider runs for its own models, plus 2 open-source runtimes you can host.
Which agent sandbox is cheapest?
There is no single answer, because the 63 price rows fall into 25 unit blocks that do not convert into each other: per vCPU-second, per physical core-second, per vCPU-hour, per Active CPU hour, per 20-minute container session, per credit and more. Within one unit the tables sort lowest first. Per vCPU-second the lowest stated rate is $0.000014 (E2B). Per vCPU-hour it is $0.01667 (Northflank Sandboxes, from the provider's general compute price list). Memory, storage, egress and plan fees are billed on top, and a low rate says nothing about startup time or reliability.
Which sandboxes use Firecracker microVMs?
E2B, Vercel Sandbox and CodeSandbox SDK state a Firecracker microVM per sandbox. The others name their own boundary. Daytona: Container class: isolated container with dedicated namespaces; VM class (Linux VM and Windows): full VM with own kernel. Modal Sandboxes: gVisor runtime; VM runtime with own Linux kernel. Runloop Devboxes: Isolated ephemeral virtual machines. Northflank Sandboxes: microVM per workload with Kata Containers or gVisor. Blaxel Sandboxes: microVMs (Mark 3.0 infrastructure). Morph Cloud: VMs (instances are running VMs started from snapshots). Beam Sandboxes: Container. Fly.io Sprites: Hardware-isolated. Cloudflare Sandboxes (Containers and Dynamic Workers): Containers: microVM per instance with own kernel and network; Dynamic Workers: separate Worker in the Workers runtime (JavaScript, Python, WebAssembly). Amazon Bedrock AgentCore Code Interpreter: Containerized environment. Azure Container Apps dynamic sessions: Hyper-V isolation. OpenAI Code Interpreter (Responses API containers): Fully sandboxed virtual machine. Anthropic code execution tool (Claude API): Linux-based container, isolated from host and other containers. Google Gemini Enterprise Agent Platform Code Execution (formerly Vertex AI Agent Engine) and Gemini API code execution do not state the technology. The feature matrix links each one to its source.
Which sandboxes block outbound network access by default?
Anthropic code execution tool (Claude API), Google Gemini Enterprise Agent Platform Code Execution (formerly Vertex AI Agent Engine) and Azure Container Apps dynamic sessions state no outbound network by default. E2B and Modal Sandboxes state outbound access is open by default, with allow and deny controls. The rest describe egress policies without stating the default, or say nothing.
Can I self-host an agent sandbox?
E2B's runtime is open source: "It is written in Go, licensed under Apache-2.0, and built by E2B." Daytona's public repository is AGPL-3.0 (GNU Affero General Public License) at tag v0.190.0, but its README says "This repository is no longer maintained." and "As of June 2026, Daytona's core development has moved to a private codebase." Among the hosted products, E2B, Daytona, Runloop Devboxes, Northflank Sandboxes, CodeSandbox SDK, Blaxel Sandboxes and Beam Sandboxes state a self-host, bring-your-own-cloud or VPC option.
How long can an agent sandbox run?
It varies by provider and plan, and several reset the clock on pause. E2B: 1 hour (Hobby), 24 hours (Pro), custom (Enterprise) of continuous runtime; pausing resets the window. Daytona: No fixed cap stated; auto-stop can be disabled so a sandbox runs indefinitely. Modal Sandboxes: Default 5 minutes, configurable up to 24 hours. Vercel Sandbox: 45 minutes (Hobby), 24 hours (Pro, Enterprise) per session; persistent sandboxes resume in new sessions. Runloop Devboxes: Default max lifetime 1 hour, configurable at creation. Northflank Sandboxes: No forced time limits stated. CodeSandbox SDK: Unlimited session length on all plans. Beam Sandboxes: Default 600-second lifetime; keep_warm_seconds=-1 for manual termination. Amazon Bedrock AgentCore Code Interpreter: Default 15 minutes, extendable up to 8 hours. Azure Container Apps dynamic sessions: Each code execution limited to 220 seconds. Anthropic code execution tool (Claude API): Containers expire 30 days after creation. Gemini API code execution: 30 seconds maximum runtime. Products not listed here do not state a maximum in the pages checked.
Are the startup times on this page measured?
No. Every startup or resume figure is what the vendor says on its own page, labelled vendor-stated. Nothing here was benchmarked, and vendors measure from different starting points (a snapshot, a template, a prewarmed pool or a sleeping sandbox).
What do inferred and Not stated mean?
Inferred marks a status the source implies but does not label: 14 products are read as generally available from published paid pricing rather than a GA badge. Not stated means the checked pages did not give the value, so the cell says so, with the note recorded for it, instead of guessing.
Last verified from each provider's own docs, pricing pages and repositories; every quote was string-matched against the page that day. Not a benchmark: prices are in the provider's own unit and change without notice, so check the linked source before you commit.
Snapshot of October 18, 2026: 20 entries (11 hosted platforms, 4 cloud primitives, 3 built-in tools, 2 open-source runtimes), 63 price rows in 25 unit blocks, 280 verified quotes.
Newsletter
Know when agent infrastructure changes
One email per week. New model releases, agent tooling and what they mean for what you build.
Keep exploring
AI agent frameworks
The code that decides what the agent runs: languages, MCP, durable execution and more.
Official MCP servers
Tools an agent can call, verified from each vendor's own docs.
LLM API cost levers
Caching, batch and service-tier discounts on the model side of an agent's bill.
Local LLM runtimes
Run the model on your own hardware next to a self-hosted sandbox.