Today I am working on a project to upgrade a TYPO3 web site from 11 to 14. This is a difficult project with APIs to external systems, git submodules and many subclasses (TYPO3 XCLASS) of the core and 3rd party extensions. I need to remove “vhs” extension and patch some extensions locally because they do not support TYPO3 14.
Normally I would do this all manually but since I am shaping my AI knowledge and practice, I do it with AI. Also AI does many routine changes very efficiently, and that speeds up the process.
We spent many hours analyzing and discussing it. AI found things, I found some things that AI missed. AI proposed solutions, I told it we can do it differently. AI checked and agreed in many cases. Human brain still matters.
Discussion took several hours. Then estimates were passed to the client, who gave a go.
I started working. Anthropic released Fable to the public again and I decided to try it. It worked well and efficiently. From time to time it stopped to ask me some questions where it was unsure. Yes, AI can be unsure if you ask it to stop if it needs to make difficult decisions. Such decisions sometimes are not right and human is still needed with its biological brain to take the final responsibility.
It all worked quite well until I hit a 5 hour limit.
AI usage limits
Generally, AI has two limits:
- 5-hour limit
- Weekly limit
You are more likely to run into the 5-hour limit if you work actively with AI.
I ran into the 5-hour limit, but my weekly limit is still OK. What I do not get is why I have a separate Fable limit. Before, it was a Sonnet limit (weaker model than the usual Opus).
Different AIs, different limits
I have both Claude Code and OpenAI codecs. Both have “x5” subscriptions. What I find interesting is that Claude Code limits end much faster than Codex'es. I use both. Generally I prefer Codex (I will write about it later), but for some tasks I use Claude code.
If Claude Code continues to reach limits so fast, I would probably migrate to Codex completely.