# Anything API workspace recovery benchmark

Measured September 9, 2026, by Notte. These results describe the tested v1 and v2 execution paths, not full conversation latency or an isolated provider effect.

## Protocol

We ran 30 paired lifecycle trials, alternating provider order. Each trial provisioned a sandbox, wrote a 5 MiB highly compressible fixture and a unique canary, verified a live lookup, suspended the environment, then recovered it and verified the file hash and canary. All measured recoveries passed verification.

V1 used Vercel Sandbox SDK 1.6.0: snapshot and stop, then create a new sandbox from that snapshot. V2 used Daytona SDK 0.201.0: stop and start the existing sandbox. The recovery timer starts after suspension or snapshot creation has finished and ends when the saved files are verified.

## Recovery results

| Metric | V1: Vercel snapshot restoration | V2: Daytona restart |
| --- | --- | --- |
| Median | 2.30 s | 1.27 s |
| p95 | 6.94 s | 1.68 s |
| Slowest | 19.69 s | 11.84 s |

The rounded medians imply approximately 45% less workspace recovery time. The recovery curve includes every attempted recovery in its denominator. Lifecycle dots and percentiles describe successful operations. With 30 observations per provider, p95 is the 29th ordered value using the nearest-rank method.

## Limits

The production v1 base image had been deleted, so Vercel used a surviving branch image with 2 CPUs and 4 GiB. Daytona used our fallback image with 1 CPU and 1 GiB. Images, regions, and warm-pool state were not matched. Recovery followed suspension immediately; the fixture was highly compressible. Timing came from a local runner and excludes model calls, agent session restoration, and browser delivery.

A separate setup pilot encountered a 90-second Daytona provisioning timeout. The successful main run does not establish production reliability. These observations do not measure time to first token, deployed-function correctness, or performance after realistic idle periods.

The article's PostHog browser-exception graph is a separate observational dataset, not part of this benchmark.