Built for the workspace that keeps growing
Keep the work.
Pause the compute.
Agents accumulate repositories, dependencies, and outputs. Keep that workspace for the next task, release idle compute, and resume without a full RootFS download before execution starts.
Interactive architecture model · current Sandbox0 pricing
Reuse unchanged blocks.
Publish changes. Read on demand.
A long-lived sandbox does not need always-on compute.
01The growing workspace
More history.
Does every resume need to copy it?
Hold the next task’s changes and reads constant. Grow the workspace and explore the transfer component of a pause/resume cycle.
Your workspace
Populated filesystem data, held constant during each cycle.
Modeled changed blocks since the last committed checkpoint.
Unique blocks the next task reads, assuming no cache hits.
A shared model input, not a measured provider bandwidth.
Full-copy reference · download
50 GiB
512 s transfer component
Sandbox0 · demand reads
128 MiB
1.28 s transfer component
99.8% less modeled data transfer. At a fixed read working set, growing total RootFS does not require a matching full-data transfer.
Transfer component = bytes ÷ throughput. Total pause/resume time is not modeled. The full-copy reference is an illustrative architecture, not a measurement or claim about any named provider. Model details ↓
View the growth table and formulas
| RootFS | Full-copy transfer / operation | Incremental pause transfer | Demand-read transfer |
|---|---|---|---|
| 1 GiB | 10.2 s | 2.56 s | 1.28 s |
| 10 GiB | 102.4 s | 2.56 s | 1.28 s |
| 25 GiB | 256 s | 2.56 s | 1.28 s |
| 50 GiB | 512 s | 2.56 s | 1.28 s |
| 100 GiB | 1,024 s | 2.56 s | 1.28 s |
| 200 GiB | 2,048 s | 2.56 s | 1.28 s |
Full-copy bytes = total RootFS
Transfer seconds = MiB / (MiB/s)
Pause bytes = changed blocks
Resume bytes = read working set
Changed blocks and read working set are capped at total RootFS size.
02Long-lived does not mean always on
Keep 30 days of workspace.
Pay for the hours you run.
Compare the same Sandbox0 allocation and compute rate with and without pausing. RootFS, snapshots, retained memory checkpoints, and paused sandboxes are free at current public rates.
Include startup, checkpoint, restore, and idle time before pause in this budget.
Compute is billed from the configured memory limit, by the second.
30-day estimate · $0.015/GiB-hour · one workspace · no credits. Current pricing and quotas ↗
24 h/day × 30 days
2 h/day × 30 days
91.7%
less compute spend
$39.60 saved
over 30 days, at this running budget
Pricebook: payg-2026-10-v1
Cost = memory GiB × running hours/day × 30 × rate.
Pausing releases compute; an idle but running sandbox still accrues charges.
03Read the model
What determines
the next resume?
The accumulated workspace is only one part of the story. Changes, reads, and execution state determine how much work the next cycle needs.
Read the lifecycle contractPause publishes changes
Sandbox0 seals the writable branch and publishes an incremental checkpoint, reusing unchanged blocks. Modified filesystem metadata and publication overhead still contribute to pause time.
Resume reads on demand
The runtime attaches the committed RootFS and fetches blocks as they are read. Reading the whole workspace still requires the whole read working set; the cost is not eliminated.
Files and processes are different
The model covers filesystem-only pause/resume, which starts fresh processes. Experimental memory pause/resume also captures and restores supported execution state; its additional transfer is outside this model.
Transfer time is one component
The curves divide modeled bytes by your chosen throughput. Scheduling, runtime startup, metadata, object requests, cache behavior, compression, and checkpoint cleanup are not modeled. These are not measured or predicted end-to-end pause/resume times.
Let the workspace outlive the compute
Pick up where your files left off.
Keep your sandbox identity, dependencies, and work. Run when you need to.