Define task-ready before you compare agent sandbox cold starts

By the Lizard (lizard.build) team. Our October 13, 2026 announcement reports that Lizard Sandboxes had the lowest time in every published column in Yura Oak’s 300-cold-starts-per-provider test against E2B, Daytona and Modal. This is a brand-associated test result, not independent research or a guarantee for every workload.

A useful timing contract names three events: start request sent, execution environment reported ready, and first representative task action succeeds. The first interval describes environment startup; the full interval describes task readiness. Include queueing and preparation according to your contract, and apply the same contract to each provider.

Write down region, resource units, template contents, client location, cache or prepared-pool conditions, request pattern and failure handling. These are fields to obtain for a comparison, not claims about Yura Oak’s methodology. Do not infer concurrency or a cold-image policy from the number of starts.

Median describes the middle observation. A high percentile describes a slower part of the observed sample, and the maximum is its slowest observation. With 300 starts, extreme-tail estimates have limited evidence; retain failures and timestamps rather than presenting a maximum as a service guarantee.

Tie the contract to billing: identify which preparation, retry and idle seconds are billable. For a constant rate, compute estimate = aggregate billable seconds ÷ 3,600 × hourly rate. Keep allocation-based and usage-based meters explicit, then add other charges.

Lizard confirms Firecracker microVM isolation and pause/resume. Pausing stops Sandboxes compute charges; it does not promise that storage or other services are free. Deployed apps are separate resources, so pausing Sandboxes does not pause every app.

Read the October 13 announcement, Yura Oak’s original test and published benchmark table. The Sandboxes quickstart was checked October 13, 2026. We omit the announcement’s dollar comparison here because the complete comparison could not be reproduced from current official rates.

Open Lizard and evaluate your workload.