An H100 GPU sandbox costs $2.95/hr on Thunder Compute and $3.95/hr on Modal and Daytona's on-demand tier. Daytona's preemptible H100 costs $2.27/hr but can be terminated without notice. All three platforms also charge for vCPU, memory and storage by the second, so headline GPU rates are only part of the bill.
This guide covers every rate that affects a GPU sandbox bill, the billing rules that change the total, and worked examples in dollars. All prices were verified in October 2026, before volume discounts or region multipliers.
Takeaways
- An H100 GPU sandbox costs $2.95/hr on Thunder Compute and $3.95/hr on Modal and Daytona on-demand.
- vCPU, memory and storage add $0.92-1.34/hr to an H100 sandbox with 8 vCPUs and 32 GiB of memory.
- Full configurations range from $3.20/hr (Daytona preemptible) to $5.29/hr (Modal).
- Preemptible GPU sandboxes cost less per hour but can be terminated without notice.
GPU Sandbox Pricing at a Glance
| Rate (per hour) | Thunder Compute | Daytona | Modal |
|---|---|---|---|
| H100 GPU | $2.95 on-demand | $3.95 on-demand, $2.27 preemptible | $3.95 (preemptible only) |
| A100 GPU 80GB | $1.95 on-demand | Not offered | $2.50 (preemptible only) |
| vCPU | $0.051 | $0.0504 | $0.0711 |
| Memory, per GiB/hr | $0.016 | $0.0162 | $0.0241 |
| Storage, per GiB/hr | $0.00012 | $0.000108 after the first 5 GiB | Volumes at $0.09 per GiB-month, first 1 TiB free each month |
| Billing increment | Per second | Per second | Per second |
| Multipliers | None | None for Linux sandboxes | 1.15x for broad region selection, 1.75x for narrow2 |
| 1 Modal's Sandbox rates list $0.00003942 per physical core per second, where one physical core equals 2 vCPU. | |||
| 2 Applies only when you pin sandboxes to a region. | |||
What a GPU Sandbox Actually Costs per Hour
A realistic H100 GPU sandbox costs $3.20-5.29/hr across these options once CPU and memory are included. The cost table prices one H100 with 8 vCPUs, 32 GiB of memory and 100 GiB of storage at list rates, with no region multiplier.
| Option | GPU | 8 vCPU | 32 GiB Memory | 100 GiB Storage | Total per Hour |
|---|---|---|---|---|---|
| Daytona preemptible | $2.27 | $0.40 | $0.52 | $0.01 | $3.20 |
| Thunder Compute | $2.95 | $0.41 | $0.51 | $0.01 | $3.88 |
| Daytona on-demand | $3.95 | $0.40 | $0.52 | $0.01 | $4.88 |
| Modal | $3.95 | $0.57 | $0.77 | $0.001 | $5.29 |
| 1 100 GiB falls within Modal's 1 TiB of free Volume storage per month. | |||||
Thunder Compute is about 20% cheaper than Daytona on-demand and 27% cheaper than Modal for this configuration. Daytona preemptible has the lowest hourly figure, but only if runs survive. An evicted run has to be paid for again.
GPU Rates for Other GPUs on Modal and Daytona
Modal and Daytona charge identical rates for every GPU but under different conditions: Daytona's GPU sandboxes are on-demand, and Modal's are preemptible.
| GPU | Daytona On-Demand | Daytona Preemptible | Modal |
|---|---|---|---|
| B300 | $7.10 | $4.08 | $7.10 |
| B200 | $6.25 | $3.59 | $6.25 |
| AMD MI355X | $5.99 | $3.44 | Not offered |
| H200 | $4.54 | $2.61 | $4.54 |
| RTX PRO 6000 | $3.03 | $1.74 | $3.03 |
| A100 40GB | Not offered | Not offered | $2.10 |
| L40S | Not offered | Not offered | $1.95 |
| RTX 5090 | $1.29 | $0.74 | Not offered |
| A10 | Not offered | Not offered | $1.10 |
| RTX 4090 | $0.99 | $0.57 | Not offered |
| L4 | Not offered | Not offered | $0.80 |
| T4 | Not offered | Not offered | $0.59 |
Billing Rules That Change Your Bill
Two billing factors beyond the headline GPU rate make the difference:
- Resources billed on top of the GPU
- Preemption
Thunder Compute, Daytona and Modal handle each rule differently.
CPU, Memory and Storage on Top of the GPU
All three platforms bill every configured resource for as long as a sandbox runs, not just the GPU. On Thunder Compute and Daytona, 8 vCPUs and 32 GiB of memory add about $0.92/hr to an H100 sandbox. On Modal, the same resources add about $1.34/hr.
Resource sizing matters as much as the GPU choice. Request only the vCPUs and memory the workload uses, since every platform bills what you configure.
Modal Sandboxes Pricing: The Sandbox Tier and Region Multipliers
Modal Sandboxes pricing has two easy-to-miss details. Sandbox CPU costs $0.00003942 per physical core per second, roughly 3x Modal's function rate, and Modal counts one physical core as 2 vCPUs. Pinning sandboxes to a region also multiplies prices by 1.15x for broad region selections and 1.75x for narrow ones (see the region selection documentation for more details).
Modal Sandboxes use Modal's standard GPU rates, and Modal GPU sandboxes are preemptible only.
Daytona Sandbox Pricing: On-Demand vs Preemptible
Daytona's preemptible GPU rate is about 42% lower than its on-demand rate for every GPU. A Daytona H100 costs $3.95/hr on-demand and $2.27/hr preemptible. Preemptible sandboxes can be terminated without notice, and evicted sandboxes remain retrievable for 24 hours.
Daytona's two tiers also behave differently when GPUs are out of stock. On-demand requests wait until capacity frees up, while preemptible requests fail immediately. vCPU, memory and storage cost the same on both Daytona tiers.
Idle Time: Timeouts and Auto-Stop Defaults
GPU sandboxes bill for the whole GPU while they exist, whether the agent keeps it busy or not. An agent whose GPU works for a tenth of a session effectively pays 10x the hourly rate per busy GPU-hour. The rest of the session goes to LLM calls and tool results.
Default timeouts cap how long an idle sandbox keeps billing:
- Thunder Compute sandboxes expire after 300 seconds by default, and the timeout can be disabled.
- Modal Sandboxes default to 5 minutes and can run up to 24 hours.
- Daytona sandboxes auto-stop after 15 idle minutes by default.
Set explicit timeouts that match the workload rather than relying on platform defaults.
Worked Example: 1,000 Coding-Agent Eval Runs
A month of 1,000 eval runs, each holding one H100 sandbox for 20 minutes, totals about 333 GPU-hours. The eval cost table uses the same 8 vCPU, 32 GiB configuration as the hourly cost table.
| Option | Cost per Hour | 1,000 Runs of 20 Minutes |
|---|---|---|
| Daytona preemptible | $3.20 | $1,0651 |
| Thunder Compute | $3.88 | $1,292 |
| Daytona on-demand | $4.88 | $1,625 |
| Modal | $5.29 | $1,7621 |
| 1 Assumes no runs are evicted; every eviction adds the cost of a rerun. | ||
Thunder Compute saves about $333/month against Daytona on-demand and $470/month against Modal for 1,000 eval runs, with no evictions to budget for.
Daytona preemptible costs less on paper, but only while evictions waste less than about 71 GPU-hours a month. Past that point, paying for lost work makes Daytona preemptible more expensive than Thunder Compute.
Daytona has no published eviction rates, so teams need to run their workloads to measure ROI. Reruns also delay results, and preemptible requests fail immediately when no spot capacity is available.
Run your evals on non-preemptible H100 sandboxes on Thunder Compute.
Cheapest Sandbox for AI Agents: Which Option Fits Your Workload?
The cheapest sandbox for AI agents depends on job length and the cost of an interruption. Sandbox pricing for AI agents only makes sense alongside those two factors.
Short, Interruptible Jobs
Daytona's preemptible tier is the cheapest option for short tasks that can be retried. When a task takes a few minutes and the harness retries on failure, an occasional eviction costs little.
Long Evals and RL Runs
Thunder Compute is the cheapest non-preemptible H100 of the three. On long evals and RL rollouts, an eviction throws away the work already done. A run guaranteed to finish at $2.95/hr can end up cheaper than a lower rate plus reruns, and the gap grows with the eviction rate.
Multi-GPU or Non-H100 Jobs
Among these three platforms, jobs that need more than one GPU, or a GPU other than the H100 or A100 80GB, currently need Daytona or Modal. Daytona supports up to 8 GPUs per sandbox and offers AMD MI355X and RTX cards. Modal covers T4 through B300.
Compare features, not just price, in our ranking of the best GPU sandboxes.
Free Credits and Startup Programs
Daytona advertises $200 in free compute, but free credits can't be used for GPU sandboxes and two new accounts we created in October 2026 each received $100.
Daytona's startup program grants $10,000 in credits on acceptance, with up to $50,000 at Daytona's discretion.
Modal's Starter plan includes $30 of compute credit per month after adding a valid payment method.
Last Thoughts on GPU Sandbox Pricing
GPU sandbox prices look similar at the headline rate and diverge once CPU, memory, preemption and idle time are added. A typical H100 sandbox costs about $3.88/hr on Thunder Compute, $4.88/hr on Daytona on-demand and $5.29/hr on Modal.
Price the full configuration you plan to run, then weigh what an interrupted run would cost before choosing a preemptible tier.
FAQ
How Much Does a GPU Sandbox Cost per Hour?
An H100 GPU sandbox (8 vCPU, 32 GiB configuration) costs $3.88/hr on Thunder Compute and $5.29/hr on Modal.
How Much Does Daytona Charge for GPU Sandboxes?
Daytona charges $3.95/hr for an on-demand H100 and $2.27/hr preemptible, plus $0.0504/vCPU-hr and $0.0162/GiB-hr of memory. Other on-demand GPUs range from $0.99/hr (RTX 4090) to $7.10/hr (B300).
How Much Do Modal Sandboxes Cost With a GPU?
Modal Sandboxes use Modal's standard GPU rates, such as $3.95/hr for an H100, and are preemptible only. Sandbox CPU and memory cost about $0.071/vCPU-hr and $0.024/GiB-hr. Region pinning adds 1.15x or 1.75x.
Do GPU Sandboxes Charge for CPU and Memory on Top of the GPU?
Yes, on Thunder Compute, Daytona and Modal. vCPU, memory and storage are billed per second alongside the GPU, adding $0.92-1.34/hr to an H100 sandbox with 8 vCPUs and 32 GiB of memory.
How Much Do 1,000 Eval Runs Cost on a GPU Sandbox?
1,000 eval runs of 20 minutes on one H100 with 8 vCPUs and 32 GiB cost about $1,292 on Thunder Compute, $1,625 on Daytona on-demand and $1,762 on Modal. Daytona preemptible costs $1,065 without evictions.
Is Spot or Preemptible GPU Pricing Worth It for Sandboxes?
Preemptible pricing pays off for short jobs that a harness can retry cheaply. For long evals and RL runs, rerunning evicted work can outweigh the discount, which makes a non-preemptible rate the safer choice.