Cloudish
5/5
apps deployed and passing
- Checks passed
- 100%
- Human steps
- 0 · 0 min
- Security & compliance
- 78/100
- Avg build
- 2m 04s
- Avg deploy
- 2m 48s
- Tokens
- 5.47M
- Agent cost
- $3.78
- Hosting, light use
- $0.55/mo
- Hosting, busy
- $39.20/mo
An open benchmark, run on GitHub Actions
AI Deploy Bench gives an AI coding agent the same five apps to build and deploy on each cloud, from a static site to a realtime chat. Then it measures what it took: whether the app worked, what a person still had to do, how secure it was, how long it took and what it cost.
Oct 13, 2026 · #8 · Full run · Opus 5.5 · medium effort · HIPAA organization · Claude Code 2.1.296
5/5
apps deployed and passing
4/5
apps deployed and passing
5/5
apps deployed and passing
4/5
apps deployed and passing
What the verifier saw on each live deployment. Open one for its checks, security review, costs and what the agent did.
The runtimes, frameworks and data stores the agent chose for each app and cloud. Open an app for its full stack, libraries and hosting.
How many of the 20 deployments used each. Names come from each app's dependencies, Dockerfile and config, and the cloud resources it created.
Across all apps, over every full run.
Share of apps whose deployment passed every required check. Higher is better.
Estimated minutes those human steps take. Lower is better.
Score of the independent security and compliance review (0-100). Higher is better.
Model API cost of the build and deploy sessions; totals across apps. Lower is better.
A fresh Claude Code session builds the app from its spec for the target cloud, without cloud credentials.
The same session gets the cloud's credentials and deploys to a public URL on its own, listing anything only a person could do.
Black-box checks test the live app over HTTP and in a browser, and screenshots record what it looks like.
A separate session scores how securely the app was built and deployed, and whether its hosting plan fits how it's used.
Everything the agent created is priced, then deleted, and the results are committed to the repository.