Skip to content
VibeCodeBench
VibeCodeBench

About

Vibe coding benchmarks for real life.

Every new AI model, the most popular tests plus the most useful everyday ones. See what it builds.

I give each AI the same prompt and put what it builds right here, so you can play it.

Say hi on X (@VibeCodeBench)

How it works

How I run them

  1. I pick a benchmark.

    Famous AI tests and useful everyday ones.

  2. Each AI gets the same prompt.

    In the apps and tools anyone can use, not the API.

  3. You play what it made.

    The file goes up untouched.

So far, agents made all 203 runs: a Claude Code or Codex agent gets the prompt, writes one file and never runs it. Those runs say so.

A run marked “auto-checked” was tried by an automated pass, not by me yet: a browser opened it, played it a little, and a model looked at the screenshots.

I keep the exact prompts private for now.

Honesty

My rules

  • Same exact prompt. Every AI gets the same text.

  • Every try stays. Broken ones too.

  • No fixes by hand. You play exactly what the AI made.

  • “1 try” means one try. If it took more, it says “try 1 of 2”.

  • No scores. You play them and decide.

  • Mistakes get a dated fix. In the log below.

Independent

Not affiliated

VibeCodeBench is an independent project. I am not affiliated with, sponsored by, or endorsed by Anthropic, OpenAI, or Google. Claude and Claude Code are trademarks of Anthropic, PBC. ChatGPT and Codex are trademarks of OpenAI. Google and Gemini are trademarks of Google LLC. Other product names belong to their owners. I use these names only to show which product made each result. It is not the Vibe Code Bench benchmark by Vals AI either.

What I fixed1 entry
  1. The first tries were missing

    13 of the 18 launch runs had been made twice. Before I published the site, it showed only the second try.

    I recovered each first try from the logs. Now both tries are shown.

    The run at max effort had 2 tries, and neither left a file.

    See Pelican on a Bike at medium effort: try 1 works, try 2 broken.

Spot a mistake? Tell me on X.