The End of Claude Code Automation

Starting June 15, Anthropic is moving automated Claude Code usage out of the normal subscription bucket.

claude -p usage will now count against a separate monthly API-style credit that is much smaller than what you were effectively getting through the normal subscription. So for heavy claude -p users it is really a large cut.

Interactive Claude Code usage still uses the normal subscription limits. This means all the tooling and automations you migh thave built around Claude Code is now worthless, they force you back into wasting your time sitting in front of the agent waiting for its output.

The product is now split in two. Use Claude manually, and you are in the normal subscription world. Use Claude from a script, and you are in a much tighter API-credit world. This effectively halfs Claude Codes usefulness for many people.

OpenAI, meanwhile, is moving in the opposite direction with Codex CLI: making the tool more broadly useful. And Codex is now at least on par with Claude Code, if not better. Personally, I think it is a good bit better.

That makes Codex the obvious landing place for developers who now have a reason to move away from Claude.

claude -p

The most important detail here is claude -p.

This is not some obscure enterprise API feature. It is Claude Code without the interactive session. You give it a prompt, it runs, you get output.

For developers, this is an extremely normal way to use a CLI tool. You put it in a shell script. You wire it into a review workflow. You use it for small one-off tasks where opening a full interactive session would be silly.

The billing line is not the product line

I understand why Anthropic wants the split. Programmatic usage is probably expensive. Some users were likely getting far more value out of subscription limits than Anthropic wanted to subsidize forever.

But from the developer side, the distinction feels artificial.

A developer using Claude Code interactively and a developer using claude -p in a script are often doing the same work. And in software development, automation is usually the point.

That is the end of Claude Code automation, which sucks because it will make many people a lot less productive.

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com