by Daytona Platforms Inc.
Sandboxed cloud computers for running AI agents and AI-written code.
Daytona suits teams building agents that execute generated code and want one vendor for CPU, Windows and GPU sandboxes. The catch is governance: core development went private in June 2026 and the public repository was archived in October. Verdict: use the hosted service, and do not plan around self-hosting.
One account covers CPU sandboxes, Windows desktops and eight GPU models, all metered by the second.
Integrations such as the OpenClaw install docs and a SOC 2 and HIPAA paper trail create some stickiness, but sandboxes are easy to swap between vendors and the archived open source core removes the self-host fallback.
Main trade-off
The open source core is frozen and archived, so anyone who wanted to self-host a maintained platform has to rely on the hosted service.
| Plan | GPU | CPU | RAM | Storage | Price |
|---|---|---|---|---|---|
| Sandbox storage (per GiB)Storage · Disk for sandboxes, priced per GiB after the first 5 GiB · United States · Europe | — | — | — | 1 GB | $0.00011/hr |
| Sandbox memory (per GiB)Cloud · Billed per second for each GiB of memory while a sandbox runs · United States · Europe | — | — | 1 GB | — | $0.02/hr |
| Sandbox vCPU (per vCPU)Cloud · Billed per second for each vCPU while a sandbox runs · United States · Europe | — | 1 vCPU | — | — | $0.05/hr |
| Windows sandbox (per vCPU)Cloud · Windows operating system rate per vCPU · United States · Europe | — | 1 vCPU | — | — | $0.09/hr |
| RTX 4090 preemptibleGPU · Preemptible capacity, can be interrupted | NVIDIA RTX 4090 | — | — | — | $0.57/hr |
| RTX 5090 preemptibleGPU · Preemptible capacity, can be interrupted | NVIDIA RTX 5090 | — | — | — | $0.74/hr |
| RTX 4090 on-demandGPU · On-demand capacity | NVIDIA RTX 4090 | — | — | — | $0.99/hr |
| RTX 5090 on-demandGPU · On-demand capacity | NVIDIA RTX 5090 | — | — | — | $1.29/hr |
| RTX PRO 6000 preemptibleGPU · Preemptible capacity, can be interrupted | NVIDIA RTX PRO 6000 | — | — | — | $1.74/hr |
| H100 preemptibleGPU · Preemptible capacity, can be interrupted | NVIDIA H100 | — | — | — | $2.27/hr |
| H200 preemptibleGPU · Preemptible capacity, can be interrupted | NVIDIA H200 | — | — | — | $2.61/hr |
| RTX PRO 6000 on-demandGPU · On-demand capacity | NVIDIA RTX PRO 6000 | — | — | — | $3.03/hr |
| MI355X preemptibleGPU · Preemptible capacity, can be interrupted | AMD MI355X | — | — | — | $3.44/hr |
| B200 preemptibleGPU · Preemptible capacity, can be interrupted | NVIDIA B200 | — | — | — | $3.59/hr |
| H100 on-demandGPU · On-demand capacity | NVIDIA H100 | — | — | — | $3.95/hr |
| B300 preemptibleGPU · Preemptible capacity, can be interrupted | NVIDIA B300 | — | — | — | $4.08/hr |
| H200 on-demandGPU · On-demand capacity | NVIDIA H200 | — | — | — | $4.54/hr |
| MI355X on-demandGPU · On-demand capacity | AMD MI355X | — | — | — | $5.99/hr |
| B200 on-demandGPU · On-demand capacity | NVIDIA B200 | — | — | — | $6.25/hr |
| B300 on-demandGPU · On-demand capacity | NVIDIA B300 | — | — | — | $6.25/hr |
Billing is per second and metered by resource. Compute is $0.0504 per vCPU-hour, memory is $0.0162 per GiB-hour, and storage is $0.000108 per GiB-hour after the first 5 GiB. Windows sandboxes add $0.0858 per vCPU-hour. GPUs are listed per hour as preemptible or on-demand: RTX 4090 $0.57 or $0.99, RTX 5090 $0.74 or $1.29, RTX PRO 6000 $1.74 or $3.03, H100 $2.27 or $3.95, H200 $2.61 or $4.54, MI355X $3.44 or $5.99, B200 $3.59 or $6.25, B300 $4.08 or $6.25. New accounts get $200 in free compute with no card required, and startups can apply for up to $50k in credits. Enterprise terms (larger limits, SSO, audit logs, bring your own compute) are quoted by sales. Rates read from daytona.io/pricing on 2026-10-04.
As stated by the provider. Ask for the current report before you rely on one.
Daytona is a hosted sandbox platform that gives AI agents isolated cloud computers, each with its own kernel, filesystem and network stack. The company says sandboxes start in under 90 milliseconds. It runs in two shared regions, United States and Europe, lists eight GPU models, and publishes SOC 2 Type I and HIPAA documents.
Per-second metering
Compute, memory and storage are each metered by the second, so a sandbox that runs for 90 seconds is billed for 90 seconds.
Fast starts (vendor claim)
The company says sandboxes reach code execution in under 90 milliseconds; Hokai has not benchmarked this figure.
Container and VM sandboxes
VM sandboxes (Linux and Windows) release disk quota when stopped or paused, while stopped container sandboxes keep it until archived.
Snapshots, builder and volumes
The platform docs list environment snapshots, a declarative image builder and volumes for starting new sandboxes from a saved state.
Eight GPU models
The pricing page lists GPUs from the RTX 4090 to the B300, each offered as preemptible or on-demand capacity.
Five SDKs plus API, CLI and MCP
Python, TypeScript, Ruby, Go and Java SDKs sit beside a REST API, a command line and an MCP server for agent tools.
| Service | Regions | Pricing | Compliance | Checked |
|---|---|---|---|---|
| US · Europe | Pay as you go | SOC 2 · HIPAA | 4 Oct | |
| -- | -- | -- | -- | |
| -- | -- | -- | -- | |
| Europe · US · Asia Pacific | Pay as you go · Hourly | ISO 27001 · GDPR | 27 Oct |
Billing is per second. Compute is $0.0504 per vCPU-hour and memory is $0.0162 per GiB-hour, while storage is $0.000108 per GiB-hour after the first 5 GiB, and Windows adds $0.0858 per vCPU-hour. By that arithmetic a 2 vCPU, 4 GiB Linux sandbox costs about $0.1656 per hour of running time. GPUs run from $0.57 per hour (RTX 4090, preemptible) to $6.25 (B300 or B200, on-demand).
There is no permanent free plan. A new account gets $200 in free compute with no credit card required, and email verification places it in Tier 1 limits. Startups can apply for up to $50k in credits through the startup program.
E2B is the closest like-for-like sandbox service on Hokai, so start there if code execution for agents is the whole job. Modal fits teams that run Python functions and models on serverless GPUs instead of long-lived sandboxes. A Hetzner server suits teams that would rather rent a machine at a flat rate and build isolation themselves.
Both meter compute by the second at close to the same vCPU rate. Per Hokai's E2B record, its Pro plan adds a monthly base fee, while Daytona's self-serve page lists usage rates only. E2B's core is Apache 2.0 and still public, whereas Daytona's repository is archived. Daytona is the one to choose when GPU and Windows sandboxes must sit in the same account.
Create an account at app.daytona.io, verify your email and copy an API key from the dashboard. Install the Python SDK or the TypeScript package @daytona/sdk, create a client with the key, call create() and then run code in the new sandbox. The vendor's own example does this in a few lines, and the $200 credit covers early experiments.
Daytona sells sandboxes: isolated cloud computers that an AI agent or a developer starts through an SDK, an API or a command line to run code. According to the company's repository, each sandbox has its own kernel, filesystem, network stack and a fixed share of CPU, memory and disk. Daytona was founded in 2023 by Ivan Burazin, Vedran Jukić and Goran Draganić, who built the Codeanywhere cloud editor starting in 2009 before moving to infrastructure for AI-generated code.
The company says a sandbox goes from code to execution in under 90 milliseconds. That is a vendor claim, and Hokai has not measured it. E2B and Modal cover similar ground, so the points that set Daytona apart are on its own pages: billing by the second for each resource, Windows sandboxes, a list of eight GPU models, and a bring-your-own-compute option for enterprise customers. Container sandboxes and virtual machine sandboxes are both offered, and the two handle stopped and paused states differently.
The typical buyer is an engineer building an agent that writes and runs code, a data tool that executes user scripts, or a coding assistant that needs a clean machine for each task. The OpenClaw install docs list Daytona as a place to run it, which makes it a natural pick for that agent's sandbox slot. Renting one machine from IDCloudHost, Hetzner or Hostinger leaves you managing it yourself, while Daytona creates and deletes machines on request. For web access rather than code execution, teams often pair a sandbox with Browserbase or Firecrawl.
Read the governance note before you plan around self-hosting. In June 2026 Daytona's core development moved to a private codebase, and the public GitHub repository, licensed under AGPL-3.0 and starred about 71,700 times, was archived on 3 October 2026 with no further updates promised. The hosted service is the maintained product. On compliance, the company's trust center post from October 2025 says SOC 2 Type I and HIPAA documents are available there. In February 2026 it raised a Series A led by FirstMark Capital, and it says the money goes to capacity and new regions.
Plans, regions, the uptime commitment and the compliance list on this page are read from Daytona Platforms Inc.'s own pages and documentation. The verdict is Hokai's reading of them, not the provider's claim.