Pay for what runs.Nothing else.

Per-second metering on Linux sandboxes and per-minute metering on iOS simulators. Paused sandboxes cost nothing.

Linux sandboxes

Metered by the second

Show rate per

$0.00000492

per physical core / second

$0.00000083175

per GiB memory / second

One physical core equals two vCPUs
Minimum allocation of 0.125 vCPU
CPU auto-scales above your size at the same rate
CPU and memory calculated independently

iOS simulators

A real device minute

$0.02/ minute

Metered while the session is active.

Browser-streamed iOS simulator
API and CLI control included
Metered per simulator minute

Migrating from another sandbox platform?

Keep core lifecycle and command code familiar while you move workloads incrementally. Unsupported provider-specific features fail explicitly.

Example adapter · ModalPython

compat_modal.py

# from modal import Sandboxfrom runcloud.compat.modal import Sandbox # Core create, exec, lookup, and terminate calls keep their familiar shape.

Compatibility matrix

One runtime. Eight migration paths.

Start with a provider-shaped import, keep the core lifecycle familiar, then validate the provider-specific features called out by each adapter.

Swipe to compare →

Sandbox provider compatibility adapters available for Run Cloud
Provider	Language	Original package	Run Cloud adapter	API shape	Migration	Guide
01Modal	Python	modal	runcloud.compat.modal	Sandbox	Core import change	Compare →
02E2B	Python	e2b	runcloud.compat.e2b	Sandbox	Core import change	Compare →
03Blaxel	Python	blaxel.core	runcloud.compat.blaxel	SandboxInstance	Core import change	Compare →
04CodeSandbox	TypeScript	@codesandbox/sdk	@run-cloud/sdk/compat/codesandbox	CodeSandbox	Core import change	Compare →
05Vercel	TypeScript	@vercel/sandbox	@run-cloud/sdk/compat/vercel	Sandbox	Core import change	Compare →
06Cloudflare	TypeScript	@cloudflare/sandbox	@run-cloud/sdk/compat/cloudflare	getSandbox	Core import change	Compare →
07Fly.io	TypeScript	Fly Machines REST API	@run-cloud/sdk/compat/fly	Machines	Core import change	Compare →
08Daytona	Python	daytona	runcloud.compat.daytona	Daytona	Core import change	Compare →

Unsupported provider-specific methods throw with a migration hint.@run-cloud/sdk · pip install runcloud-sdk

Linux compute

Compare the rates.

Adjust concurrency and monthly runtime to compare normalized Linux compute estimates across providers.

Concurrent sandboxes50Runtime per sandbox / month240 hr / 10 days

Region selection

Broad and exact placement apply only published region-selection premiums. Providers without a published multiplier remain unchanged.

Normalized to 2 vCPU / 4 GiB. Estimates apply each provider's published usage rates or nearest matching tier to the same runtime.

Run Cloud1×

Per-second CPU + memory

$356.27

Fly.io≈3.3×

Performance-2x / 4 GB regional rate

$1,172.40

CodeSandbox≈5.1×

On-demand VM credit rate

$1,800.00

E2B≈5.6×

Per-second CPU + memory

$1,987.20

Daytona≈5.6×

Per-second CPU + memory

$1,987.20

Blaxel≈5.6×

Active compute by allocated memory

$1,987.20

Cloudflare≈6.1×

Containers at 100% active CPU

$2,160.00

Modal≈8.0×

Sandbox rate at 3× base CPU + memory

$2,848.61

Vercel≈11.5×

Sandbox at 100% active CPU

$4,089.60

Illustrative usage estimate only. Plan fees, included credits, storage, network, and creation charges are excluded. Active-CPU providers assume 100% utilization; lower utilization can reduce their actual bill.

Competitor prices reviewed July 26, 2026 from official sources: Modal, Modal region selection, E2B, Blaxel, Box, CodeSandbox, Vercel, Cloudflare, Fly.io, Daytona. Prices and billing models can change; verify the linked source before making purchasing decisions.

Real sandbox benchmarks

Price is only useful with performance.

The open-source StarSling suite runs real repositories alongside CPU, memory, disk, database, and network tests. The common target is 4 vCPU, 8 GiB RAM, and 40 GB disk.

Better-Auth

202.3s

Third-fastest measured total

PostgreSQL RW

18,357 TPS

Highest measured throughput

Dax benchmark

#1*

Unofficial public-script result

Real development workloads

Lower is better. Totals sum the task medians that produced portable results.

4 vCPU · 8 GiB · 40 GB

Real development workload benchmark results
Metric	Run Cloud	Daytona VM	Blaxel	Modal VM	Modal gVisor	Novita	Box*	E2B
Better-Auth measured totalseconds · lower ↓	202.34	160.58	167.83	214.92	442.58	225.18	357.50	286.99
Mastra measured subtotalseconds · lower ↓	231.82	194.99	199.93	245.75	445.58	282.00	465.94	—
OpenClaw measured subtotalseconds · lower ↓	57.29	48.98	48.66	58.18	115.40	68.51	106.73	—

Normalized active hourly cost

Run Cloud uses today's rates. Competitor figures use the benchmark-recorded rate, except Box and Blaxel, which are calculated from their current published billing models.

Normalized active hourly benchmark cost
Metric	Run Cloud	Daytona VM	Blaxel	Modal VM	Modal gVisor	Novita	Box*	E2B
Active hourly cost at benchmark shapeUSD/hour · lower ↓	$0.05938	$0.2502	$0.3312	$0.7612	$0.7612	$0.2333	$0.036*	$0.3312

Run Cloud: (2 physical cores × $0.00000492 + 8 GiB × $0.00000083175) × 3,600 = $0.0593784/hour. One physical core equals two vCPUs.

* Box: $20 ÷ 2,000,000 included active seconds × 3,600 = $0.036/hour. Using Box requires a minimum $20 monthly plan.

Blaxel: 8 GB × $0.0000115 × 3,600 = $0.3312 per fully active hour. Suspended time is not active CPU; snapshot storage at $0.20/GB-month and images at $0.045/GB-month are excluded.

View all 44 benchmark measurementsCPU, memory, disk, databases, network, and task-level timings

CPU and memory

Compiler and memory-throughput signals. Higher is better.

CPU and memory benchmark results
Metric	Run Cloud	Daytona VM	Blaxel	Modal VM	Modal gVisor	Novita	Box*	E2B
Node.js web toolingruns/s · higher ↑	18.19	18.60	19.80	15.07	9.56	13.52	8.60	11.24
STREAM CopyMB/s · higher ↑	44,496	191,557	102,067	81,873	83,563	57,121	73,625	75,238
STREAM ScaleMB/s · higher ↑	30,178	170,518	94,974	48,134	54,681	47,124	49,563	43,027
STREAM AddMB/s · higher ↑	32,984	177,524	101,989	55,278	64,919	49,254	53,307	48,461
STREAM TriadMB/s · higher ↑	33,021	178,193	100,638	55,851	67,029	49,300	52,869	48,703

Disk and filesystem

fio uses O_DIRECT. Higher is better for every row.

Disk and filesystem benchmark results
Metric	Run Cloud	Daytona VM	Blaxel	Modal VM	Modal gVisor	Novita	Box*	E2B
Random read, 4 KiBIOPS · higher ↑	159k	245k	249k	230k	31k	79k	43k	48k
Random read, 4 KiBMB/s · higher ↑	620	956	970	896	121	310	169	187
Random write, 4 KiBIOPS · higher ↑	169k	219k	218k	210k	26k	77k	39k	49k
Random write, 4 KiBMB/s · higher ↑	658	855	850	818	103	302	154	193
Sequential read, 1 MiBIOPS · higher ↑	4,624	8,352	8,644	2,018	19,950	9,492	1,778	599
Sequential read, 1 MiBMB/s · higher ↑	4,620	8,354	8,631	2,019	—	8,545	1,779	601
Sequential write, 1 MiBIOPS · higher ↑	5,403	3,973	5,874	3,127	2,888	5,405	2,092	599
Sequential write, 1 MiBMB/s · higher ↑	5,405	3,974	5,875	3,129	2,890	5,406	2,093	601
Hardlink throughputops/s · higher ↑	5.71	25.83	19.49	8.09	3.02	15.88	3.38	1.40

System and databases

Arrows show whether a higher or lower result is better.

System and databases benchmark results
Metric	Run Cloud	Daytona VM	Blaxel	Modal VM	Modal gVisor	Novita	Box*	E2B
PyBenchms · lower ↓	483.5	411.5	486	815.5	897	677	1,138	805.5
Git common operationsseconds · lower ↓	39.61	39.14	45.06	63.07	84.66	51.28	70.36	65.15
SQLite Speedtestseconds · lower ↓	78.64	32.18	38.95	62.89	429.69	58.63	157.06	71.80
PostgreSQL read-onlyTPS · higher ↑	217,940	279,897	289,700	192,875	10,991	184,917	99,331	182,075
PostgreSQL RO latencyms · lower ↓	0.230	0.179	0.173	0.259	4.549	0.272	0.504	0.275
PostgreSQL read/writeTPS · higher ↑	18,357	15,737	17,786	13,710	1,905	14,095	6,830	12,635
PostgreSQL RW latencyms · lower ↓	2.724	3.177	2.815	3.647	26.249	3.550	7.323	3.958

Network

Mbit/s. Higher is better. A dash is a retained benchmark gap.

Network benchmark results
Metric	Run Cloud	Daytona VM	Blaxel	Modal VM	Modal gVisor	Novita	Box*	E2B
Loopback TCP, 1 streamMbit/s · higher ↑	57,315	78,354	109,237	15,201	15,574	149,130	22,886	56,874
Loopback TCP, 10 streamsMbit/s · higher ↑	57,765	98,213	105,696	15,405	13,946	127,395	26,428	52,622
Loopback UDP, 10G targetMbit/s · higher ↑	9,999	9,999	9,999	9,999	169	9,999	10,000	9,999
WAN downloadMbit/s · higher ↑	746	3,421	1,054	1,415	7,936	4,401	1,158	2,477
WAN uploadMbit/s · higher ↑	909	4,746	1,622	5,095	158	5,589	5,732	1,909

Better-Auth task breakdown

Seconds. Lower is better.

Better-Auth task breakdown benchmark results
Metric	Run Cloud	Daytona VM	Blaxel	Modal VM	Modal gVisor	Novita	Box*	E2B
Git cloneseconds · lower ↓	1.94	1.49	0.82	0.79	2.44	1.92	2.53	1.49
Cold installseconds · lower ↓	11.74	10.15	10.82	17.25	31.27	12.74	19.72	17.86
Buildseconds · lower ↓	73.30	57.72	59.35	76.19	142.89	83.36	125.52	97.95
Biomeseconds · lower ↓	3.49	3.11	3.31	4.09	10.93	4.24	6.21	5.30
Knipseconds · lower ↓	11.62	10.57	10.44	13.47	29.66	13.89	25.08	18.90
Formatseconds · lower ↓	3.26	2.86	3.09	3.75	7.41	3.98	6.59	5.22
Spellseconds · lower ↓	7.90	7.28	7.19	9.13	16.41	9.18	18.87	13.36
Type lintseconds · lower ↓	36.10	25.40	26.80	35.72	109.32	39.70	58.22	50.50
Package lintseconds · lower ↓	2.52	2.41	2.57	3.24	11.02	3.27	4.85	4.24
Typecheckseconds · lower ↓	50.47	39.59	43.45	51.30	81.22	52.91	89.92	72.17

Mastra and OpenClaw tasks

Seconds. Lower is better. Mastra test_core and five OpenClaw tasks are excluded because they produced benchmark gaps.

Mastra and OpenClaw tasks benchmark results
Metric	Run Cloud	Daytona VM	Blaxel	Modal VM	Modal gVisor	Novita	Box*	E2B
Mastra · Git cloneseconds · lower ↓	6.70	2.84	1.99	2.96	6.36	3.34	7.65	—
Mastra · cold installseconds · lower ↓	33.31	25.39	27.20	33.93	66.79	41.43	63.80	—
Mastra · format lintseconds · lower ↓	104.79	93.56	95.25	115.87	199.17	130.90	222.18	—
Mastra · build coreseconds · lower ↓	87.02	73.21	75.49	92.99	173.26	106.32	172.31	—
OpenClaw · Git cloneseconds · lower ↓	7.08	3.19	2.64	3.60	9.39	4.51	11.51	—
OpenClaw · cold installseconds · lower ↓	6.69	5.95	5.44	6.60	14.86	8.15	11.91	—
OpenClaw · typecheckseconds · lower ↓	43.52	39.84	40.59	47.99	91.16	55.85	83.32	—

Run Cloud and the published provider results use replicated medians. Box is a complete single-machine sweep on one persistent out-of-the-box VM, so its figures are directional rather than publication-grade. A dash is a retained benchmark gap, never a score of zero.

Mastra's test_core hit the same JavaScript heap limit across providers. OpenClaw totals include clone, install, and typecheck; five other declared tasks produced gaps. E2B could not run Mastra or OpenClaw with the available free disk.

* Run Cloud's Dax result was produced by us using the public benchmark script and would rank first, but it is not yet included in ComputeSDK's published provider runs.

Methodology and the original comparison: benchmark repository · StarSling analysis · ComputeSDK Dax benchmark

The landscape

A directional view of the sandbox providers covered by the Run Cloud compatibility layer. Change the dimension to see how each product is positioned.

affordableexpensivecomplex setupone command

Run Cloud
Cloudflare
Fly.io
Daytona
Blaxel
E2B
CodeSandbox
Modal
Vercel

Directional product positioning based on publicly documented capabilities reviewed July 2026. Positions are qualitative, not benchmark scores, and products change over time.

Worked examples

Tiny numbers. Predictable totals.

Illustrative Linux costs assume the listed allocation runs continuously for one hour. These price the reserved size; CPU used above it bills at the same rate as a separate CPU (burst) line.

Small agent

$0.00593

per running hour

CPU: 0.25 core
Memory: 0.5 GiB

Standard agent

$0.02370

per running hour

CPU: 1 core
Memory: 2 GiB

Parallel build

$0.04740

per running hour

CPU: 2 cores
Memory: 4 GiB

Pricing FAQ

The practical details.

How the meter works, and what the rates do and do not include.

How is Linux sandbox usage calculated?

Linux sandbox compute is metered by the second. CPU costs $0.00000492 per physical core-second and memory costs $0.00000083175 per GiB-second. CPU and memory usage are metered independently.

What does one physical core mean?

One physical CPU core is equivalent to two vCPUs. Linux sandboxes have a minimum CPU allocation of 0.125 vCPU (0.0625 physical cores).

What happens when a sandbox needs more CPU than its size?

Sandboxes auto-scale CPU. The size you pick is a billing floor: reserved CPU is always charged while the sandbox runs, and any CPU consumed above it bills at the same per core-second rate, shown as its own CPU (burst) line on the invoice. Memory does not auto-scale; it is a hard reservation.

How much does an iOS simulator cost?

Simulator sessions cost $0.02 per active minute. Metering runs while the session is active and stops when it is released or times out.

Will Linux sandbox rates change?

Only downward. The rates shown here are Run Cloud launch pricing and we will not raise them. As our costs fall, we pass the savings on. Any decrease is published before it affects billing.

Start in beta. Build for real.

Try Linux sandboxes, open an iOS simulator, and ship your first Xcode build without maintaining the fleet.

Start building Explore the platform