As long as there are no actual logical constraints for what the generated code must do you will get all sorts of weird behaviors like deleting the entire database to make the tests pass or just making all the tests vacuously passing to avoid failures in CI. Social media is full of examples like this one where the code generator does not actually follow the intent of the programmer b/c there is no way to specify logically that tests should not be vacuously true : https://x.com/Sauers_/status/1964354357635285391. That example is from Claude Code but there are plenty of failures for all the other vibe coding tools.

Bundling and distribution. OpenAI has more paid subscribers than Anthropic. I started using Codex over Claude because Codex is included in my subscription.

It’s the classic Microsoft Teams vs Slack debate.