Benchmarks

cf-orbs against Amp-hosted orbs, and the Docker, E2B and Vercel Sandbox providers from Amp's examples. Same prompt, same project. Orb time only, lower is better.

Every run counted, failures included, no best-of. Each cell is the median (nearest rank). cf-orbs ran at 4 vCPU / 12 GiB; Amp-hosted at a1.small (2 CPU, 4 GiB). The size of each column is under its name.