Every coding agent needs somewhere to execute code, and the sandbox market has exploded into a dozen vendors with four incompatible billing models and cold start claims measured under conditions nobody publishes. The real decision comes down to four properties: cold start under concurrency, filesystem persistence between turns, egress policy, and idle billing. Vendor claims are not comparable. Daytona advertises sub-90ms sandbox creation. E2B is cited around 150ms. Modal claims sub-second cold starts for pre-cached containers. None state concurrency, region, image size, or whether the clock stops at API acknowledgment or first executed command.
The most useful public dataset is ComputeSDK’s open-source sandbox leaderboard, which measures Time to Interactive (TTI) with 100 iterations per provider, launched concurrently in a single burst from a 4 vCPU host in Northern Virginia. The August 21, 2026 run reveals three things more important than ranking. Burst is not the same test as sequential: Daytona posted the fastest median in the field but completed only 37 of 100 attempts, meaning a median you only reach on a third of your calls is a capacity number, not a latency number. Tail latency is what to design against: Runloop’s P95 is 3.3x Modal’s despite their medians being 10ms apart. Cloudflare’s Sandbox SDK sits on Cloudflare Containers, architecturally heavier than resuming a pre-warmed Firecracker VM, and its 5s medians reflect that.
Pricing normalized per second shows Modal prices per physical core, which it defines as 2 vCPU. Modal’s sandbox tier runs roughly 3x its standard Function rate at $0.00003942 per core-second versus $0.0000131, and region selection adds 1.5 to 1.75x on top. Daytona’s GPU rates are widely misquoted: third-party tables list H100 at $3.95 per hour, but the live pricing page shows $2.27 per hour for on-demand H100 and $2.61 for H200. Comparison tables in this category go stale within a quarter.
The practical advice is to stop trusting vendor numbers and measure the task your agent actually runs. A useful harness runs the same unit of work everywhere—install pandas, read a CSV, plot it, return a PNG—and times four checkpoints separately: create, TTI, dependency install, and task completion. Report TTI and task separately because vendors optimize the first and readers care about the second. Pin the region, pin the image, and publish both sequential and concurrent series. This mirrors what happened with serverless function benchmarks a decade ago, where the gap between marketing cold starts and real-world concurrency behavior only became visible once independent measurement forced vendors to be honest about their numbers.
