Sandbox comparison · Checked October 21, 2026
Agent37 vs CodeSandbox SDK
Published running-cost comparison
95% less
$4.76 with Agent37 vs $108.48 with CodeSandbox SDK
$103.72 less per sandbox per 730-hour month.
730 running hours at published rates, before credits and plan fees. Agent37: shared 2 vCPU, 4 GB RAM and 4 GB disk. CodeSandbox SDK: Nano. Full CPU consumption where usage-billed; hardware, storage and session limits differ.
CodeSandbox SDK is now documented as a legacy product by Together AI. Its development environments support VM checkpoints, cloning and source-control integration.
Lifecycle performance, side by side
Agent37 figures use our Performance dashboard telemetry from October 22, 2026 at 00:50 UTC to October 22, 2026 at 01:15 UTC, after the latest speed improvements. This is an early snapshot; small samples can vary. CodeSandbox SDK figures are attributed to public benchmarks or vendor sources. Each cell states what was timed: API calls, host restores and first-command latency are different measurements, so this is not a normalized speed ranking.
| Operation | Agent37 | CodeSandbox SDK |
|---|---|---|
| Create | 1.29 s median Measured · n = 37 Successful create API duration, from entering API authentication through the response. The sandbox has started; this does not wait for tenant application readiness or the first command. Mixed shapes, images, host regions and runtimes. | 2.7 s p95 Provider-published timingMay 2025 launch benchmark for a fresh CodeSandbox-based VM. This describes the legacy SDK product, not the current Together Sandbox SDK. |
| Resume / wake | 176 ms median Measured · n = 38 Host-timed restoration of a same-host gVisor memory checkpoint, including local startup housekeeping. Excludes the client round trip and subsequent command. Host reports are not filtered by workspace; cold boots and no-ops are excluded. | Within 1 s Provider-published timingThe legacy SDK README claims snapshot restoration within one second. This is restoration of saved VM state, not a fresh image build. |
| Sleep / suspend | Supported; timing not collected Auto-sleep can checkpoint memory. The current telemetry does not retain checkpoint duration. Stop timing below measures a different operation. | No numeric timing in reviewed sources. |
| Run a command | 461 ms median Measured · n = 811 Successful exec API duration, including the tenant command runtime. Commands vary across requests. This is not a controlled echo benchmark. Excluding host-reported command time: 338 ms median API/network overhead. | No numeric timing in reviewed sources. |
| Stop | 2.53 s median Measured · n = 2 Successful stop API duration. Stop discards memory state, preserves durable files and waits for the operation response. Includes requests to instances already stopped; this is not memory-preserving sleep latency. | No numeric timing in reviewed sources. |
| Restart | 1.26 s median Measured · n = 4 Successful restart API duration, including control-plane work and the restart operation. Mixed shapes, images, host regions and runtimes. | No numeric timing in reviewed sources. |
| Delete | 429 ms median Measured · n = 32 Successful delete API response duration. The response acknowledges deletion and ends billing; runtime stop, archive backup and physical teardown continue asynchronously. This is not time to completed cleanup. | No numeric timing in reviewed sources. |
Median is p50. Request counts are timing observations, not customers. Shapes, images and commands vary. Missing timing data does not mean a feature is unsupported.
Other Agent37 wake paths and measurement notes
Memory restore, VM
No observations in this window
n = 0. Host-timed restoration of a same-host VM memory checkpoint through the guest resume RPC and local housekeeping. Excludes the client round trip and subsequent command. Host reports are not filtered by workspace; cold boots and no-ops are excluded.
Wake by cold boot, default sandbox
No observations in this window
n = 0. Host-timed wake that cold-boots instead of restoring memory. Durable files remain, while live process state is not restored. Excludes the client round trip; host reports are not filtered by workspace.
Wake from cold storage
No observations in this window
n = 0. Control-plane-timed recovery of a sleeping instance from cold storage onto a host, including data recovery and boot. Process memory is not restored. Separate boundary from same-host memory restoration.
Wake after host relocation
No observations in this window
n = 0. Control-plane-timed wake that moves the sleeping instance to another host, including data transfer and boot. Process memory is not restored. Separate boundary from same-host memory restoration.
- This is an early production snapshot after the lifecycle performance rollout, not an SLA or a controlled comparison. The explicit window excludes traffic before the rollout. Shapes, image sizes, host regions and runtimes vary; low sample counts are less stable than a full-day sample.
- API metrics select successful 2xx responses, excluding the configured integration workspace and recognized automated-test workspace prefixes. API timers run from entering authentication through response creation, including downstream work but excluding the caller internet round trip.
- Exec includes arbitrary tenant command runtime. Exec overhead subtracts the host-reported command time; it is not equivalent to a full end-to-end command latency benchmark.
- Default and VM memory-restore timers run on the host. They exclude the client round trip and the subsequent application request or command. Host wake telemetry is not filtered by workspace. Counts are timing reports, not unique customers or guaranteed unique wake events.
- Cold boots, cold-storage recovery and host-relocation wakes have different paths from same-host memory restoration and are reported separately. Already-running wake no-ops are excluded.
- Start API timing is omitted because the aggregate includes already-running no-ops, which would make its median misleading as a resume or cold-start measurement.
- No sleep/checkpoint-duration aggregate is collected by the existing Performance helper. Stop is not a substitute for memory-preserving sleep.
- Observations are sorted and zero-based index floor(p * n) selects p50 and p90, with p = 0.5 and 0.9. The window includes its start and excludes its end. Counts represent operation observations, not customers. No observations is represented by null timing and zero samples, never zero latency or a carried-forward result.
- Only aggregate timings are published. No customer, workspace, instance, host, credential or raw-log data is included.
What 730 running hours cost
Reference workload: 2 vCPU and 4 GB RAM. USD list rates before credits, taxes and plan fees. This is aggregate runtime, not a promise of one uninterrupted session or equal hardware performance. GB and GiB follow each provider’s units. Full CPU consumption is assumed for usage-based CPU billing; idle-heavy workloads can cost less.
Agent37 Standard
$4.76
2 vCPU · 4 GB RAM · 4 GB persistent disk
Shared compute for bursty agent work. Sustained heavy CPU use may be limited on busy hosts. Choose Performance for reserved cores, at 4x standard compute.
Agent37 pricing and calculatorCodeSandbox SDK
$108.48
Nano/Micro/Small: 10 / 20 / 40 VM credits per hour at $0.01486 per credit
Quoted size: Nano
Legacy CodeSandbox SDK rates; Together recommends its successor for new integrations. Runtime rounds up to the minute. Plan fees and included credits excluded; Build allows 10 concurrent VMs. Hibernation stops compute billing.
CodeSandbox SDK pricing source100-workspace cost estimate
$10,371.80 less per month.
$476.00 on Agent37 vs $10,847.80 on CodeSandbox SDK.
Illustrative 100-workspace total at 730 running hours each, using the same rates and assumptions above. Totals use unrounded rates; volume discounts and account limits are not included.
Agent37 includes retained disk in this figure; competitor storage is excluded unless stated. Free allowances, minimum commitments, resource allocation, region and session limits can change your actual bill. Read the comparison method.
Why Agent37
Where Agent37 fits
Choose Agent37 for a lasting machine with native SSH, ready-made agent templates and inexpensive compute. You can build the agent experience around a persistent workspace without adopting a legacy SDK product.
Start with a ready-made agent template or publish your own image. Keep files, expose apps and manage the machine through one API. Models and managed tools are optional and billed separately.
Requirements that could favor CodeSandbox SDK
An existing application may need CodeSandbox's exact checkpoint and cloning behavior during migration. For new work, verify the successor Together Sandbox's access and terms; legacy prices are not a quote for that product.
For a persistent, bursty Linux agent without those requirements, Agent37 is our recommendation at the prices shown.
Beyond the hourly rate
| Factor | Agent37 | CodeSandbox SDK |
|---|---|---|
| Persistence and lifetime | Keep the same instance and persistent data between tasks, with no fixed session deadline. Auto-sleep and stop reduce compute charges; retained disk still bills. | VM checkpointing and restoration are supported, alongside cloning and hibernation workflows. Confirm current retention and plan allowances before sizing a deployment. |
| Isolation and runtime | Standard sandboxes use gVisor. Choose a VM for its own Linux kernel and Docker support, at 2.5x standard compute. Standard and ordinary VM instances share host resources. | CodeSandbox microVM infrastructure. Custom environments can be built from Dockerfiles. |
| Billing | Compute bills by the minute while awake, including CPU-idle time. Stopped or sleeping entry disk costs $0.36 per 730-hour month. No plan fee. | Legacy CodeSandbox SDK rates; Together recommends its successor for new integrations. Runtime rounds up to the minute. Plan fees and included credits excluded; Build allows 10 concurrent VMs. Hibernation stops compute billing. |
Provider startup claims and our measurements use different boundaries. A resumed snapshot, a prewarmed session and a fresh machine are different operations. Compare all providers and timing sources.
Sources and verification
Checked October 21, 2026. Prices and features can change. Provider descriptions come from the official pages below; workload recommendations are our assessment.
Try your workload on Agent37.
Start your next agent on Agent37. Persistent files, your choice of agent and a standard sandbox from $4.76 per 730 running hours. Measure it against the latency and resource needs of your application.
Eligible new accounts. Add a card to unlock credit; no charge until you top up. Model and tool usage is separate.