cloudishai/ai-deploy-bench is a benchmark that reproduces the process in which AI coding agents build web apps and deploy them across multiple clouds, measuring not only whether the deployment succeeds but also the time taken, token usage, cost, the degree of human intervention, and the security status. The clouds covered are Cloudish, DigitalOcean, Render, and Vercel. The latest commit in this repository, 2ccabd6 (2026-10-15), contains not new measurements but changes on the site side that publish those results. Across 17 files, 613 lines were added and 19 lines were deleted.
What was added is the coexistence of human-readable HTML and the Markdown that AI agents read. Every page now comes with an index.md version in the same folder, and nginx returns the page address itself, with a markdown content type, for requests that send Accept: text/markdown. All links are fixed as absolute paths, so the form is preserved wherever it is fetched. /llms.txt indexes the site, and /llms-full.txt gathers the overview and methodology, recent run records, and the per-cloud pages into a single file. Each page's head contains a canonical URL, alternate links pointing to the Markdown and JSON, and schema.org Dataset JSON-LD, and robots.txt allows crawlers while pointing to llms.txt and the sitemap. Sitemap entries carry a lastmod date. A CI job was added that builds the nginx image directly to check the content types, Markdown negotiation, the Vary header, and the CSP.
However, what this commit touched is not the benchmark's measurements or run records but the way results are published. Whether any new success rate or cost figure was added cannot be known from this commit alone; only the way the already-accumulated results are laid out across pages and data files has changed. Among the individual result items, some are copied as-is from past runs, so not every item is the result of a newly executed run. No license or star count is given in the provided information. If you are building an AI agent that reads and uses the site, you can say it now has a form you can refer to, but to gauge the reliability of the benchmark itself, you need to read the run results and the methodology separately.