Comparisons
Modal, E2B or Daytona? Your agents don't have to pick one.
In six weeks Modal passed a $1B run rate, Daytona raised $60M at a $1B valuation, and E2B shipped two security features. All three sell computers for AI agents. Each is built for a different kind of work, so the right answer is to use all three.
Agents need computers. A coding agent needs somewhere to run tests, a browser agent needs a screen, and a model fine-tune needs a GPU. Three companies have become the default places to rent those computers: Modal, E2B and Daytona. This fall each of them shipped a lot. Here is what changed, what each is good at, and what the same hour costs on each.
Modal: the GPU cloud that also runs a billion sandboxes
- $750M Series D, October 14. $15.75B valuation, more than $1B in annualized revenue, about $300M of it from sandboxes, and nearly 3 billion sandboxes created. (Modal)
- VM Sandboxes are generally available, October. Set
runtime="vm"and a sandbox becomes a full Linux VM: Docker, Docker-in-Docker, desktop apps, mobile simulators. Modal says cold starts take under a second. (Modal) - Clusters and a Training API, October. Multi-node GPU clusters are generally available, and a new API handles model training. (Clusters, Training API)
- Sandbox Sidecars put a trust boundary next to each sandbox. (Modal)
Best for: anything that needs a GPU, such as inference, training and RL, plus huge bursts of parallel sandboxes for evals.
E2B: secure code execution, now with secrets handled for you
- Workload Identity, November 3. Each sandbox gets a signed identity, so your backend can check who is calling without handing out long-lived keys. (E2B)
- Secrets, October 7. Credentials stay outside the sandbox and are added to outbound HTTPS requests, so the agent never sees the key. (E2B)
- E2B Embed, September 30. An Apache-2.0, single-machine bundle you run in your own environment, aimed at government and regulated teams. (E2B)
- Agents can sign up on their own through Stripe Projects (June), and E2B is a built-in provider in the OpenAI Agents SDK (April). (E2B)
Best for: running untrusted code, code interpreters, data analysis and browser agents on a Linux desktop, especially when security review is the blocker.
Daytona: a whole computer per agent, any operating system
- $60M Series B at $1B, October 20. Total funding is $91M, and Daytona reports about 25× revenue growth since January. (Daytona)
- Sub-90 ms provisioning, October 29. Daytona describes microVMs started from pre-warmed snapshots, and a runtime you can deploy on your own clusters. (Daytona)
- Computer use, August. Linux desktop VMs from a pre-booted pool, streamed to a browser. Daytona says it is the only provider with Linux, Windows, macOS and GPU sandboxes on one platform. (Daytona)
- Cursor self-hosted machines run on Daytona (September). Daytona also went closed source in June. (Daytona)
Best for: long-lived computers that pause, fork and rewind, Windows and Mac desktops, and cheap preemptible GPUs.
The same hour on each
We priced one sandbox with 2 vCPU and 4 GiB for one hour, plus one H100 hour, from each pricing page in November 2026:
| 2 vCPU · 4 GiB, 1 h | H100, 1 h | Free to start | |
|---|---|---|---|
| E2B | $0.17 | no GPUs | $100, once |
| Daytona | $0.17 | $2.27 preemptible | $200, once |
| Modal | $0.24 | $3.95 | $30 every month |
The CPU hour costs the same on E2B and Daytona. Modal charges sandboxes 3× its list rate, so a plain sandbox there costs about 40% more. For GPUs it flips: Modal has the broadest GPU platform, and Daytona's preemptible H100 is cheaper but can be taken back. No row wins every column.
Math: E2B $0.000014 per vCPU-second and $0.0000045 per GiB-second; Daytona $0.0504 per vCPU-hour and $0.0162 per GiB-hour; Modal sandboxes $0.00003942 per core-second (1 core = 2 vCPU) and $0.00000667 per GiB-second, H100 $0.001097 per second. Sources: e2b.dev/pricing, daytona.io/pricing, modal.com/pricing.
Why you don't have to choose
Every comparison post ends with "pick the one that fits." That made sense when an agent did one kind of work. It doesn't anymore. One task from one agent might be: read a spreadsheet, fill a form in a browser, run a script, then caption a video. That is a desktop, a sandbox and a GPU, three different best answers in a single task.
Pick one provider and you overpay for the work it isn't built for, wait on cold starts it can't avoid, and stop when it goes down.
- Each job on the computer built for it. GPU work on Modal, untrusted code on E2B, desktops on Daytona, private files on your own Mac.
- Price. The same CPU hour ranges from $0.17 to $0.24 across the three. Across thousands of agent runs, routing decides the bill.
- Uptime. One provider is one point of failure. With two, an outage means the work moves instead of stopping.
- Free credits. $30 a month plus $100 plus $200 in starting credit is more than any one of them gives you.
One roof: Cerver
Cerver is one API over all of them, plus your own machines. You connect your own provider accounts, each provider bills you directly, and Cerver adds nothing to the compute. Your agent says what a task needs, and Cerver starts the computer that fits. It can move a running session to another computer mid-task: in our test an E2B session was ready in 26 seconds and moved to Vercel in 28.
What runs today: E2B, Vercel and your own machines. The Modal and Daytona connections are built and in testing; we will only mark them live once they pass the same end-to-end run.
Use every provider you already have, through one API.
Connect your E2B and Vercel accounts and your own machines, and let your agents pick the computer for each job. Plans start at $20 a month, on your own provider keys.