← All benchmarks

Web Clone Bench

Code · 2026-10-05

Plato's Web Clone Bench v1: an agent is given a running reference build of one of nine web apps (Gmail, Figma, Google Sheets, Slack, DocuSign, Jira, Shopify, Craigslist and Zendesk) and an empty directory, with no written specification or starting code, and has up to 8 hours on an 8-CPU machine to rebuild the product from scratch in any stack; shipping or proxying the reference's own build output disqualifies the run. User stories written by domain experts become end-to-end verifiers that replay each workflow against the clone and compare its visual and functional behavior and state changes with the reference. Each app gets a full score in percent, and the benchmark score is the mean full score over the nine apps. Plato runs each model in its own vendor's coding harness at each reasoning effort it tests.

Top models (higher is better)