I ping pong back and forth between claude code and codex.
In my experience (very subjective, obviously) for backend/"logical" tasks Codex seems to outperform Claude.
For front-end/UX related tasks Claude wins easily.
Overall, Claude does seem to be a little better in other areas too.
Codex's biggest advantage in my personal opinion, however, is usage. I think maybe once in several months did I even get close to hitting my limit with the $20 plan.
With Claude, however, I feel like I can sneeze and half my weekly usage is gone. Same $20 price tag.
That's been my experience, I'm sure it differs user to user though.
I've gone fully custom and would never look back. All of the popular tools lack the experience I'm after, generally lacking the advanced customization I have found makes weaker models more capable. I'm also moving towards open weight models.
The underlying motivation is to avoid Big Ai as much as possible. Incentives are not aligned.
Good prompts and tools. When I first went down this path, I grabbed the Claude Code prompts and adapted them. Doing so made gemini-3-flash as good as pro for 80%+ of my normal tasks. I became spoiled with the turnaround time (much faster) and now become impatient with the mega models.
I can also do things like out the AGENTS.md contents in the system prompt and change file read to put contents there as well. Thai avoids duplicates in the messages when they fail at editing and decide to reread the file
I ping pong back and forth between claude code and codex.
In my experience (very subjective, obviously) for backend/"logical" tasks Codex seems to outperform Claude.
For front-end/UX related tasks Claude wins easily.
Overall, Claude does seem to be a little better in other areas too.
Codex's biggest advantage in my personal opinion, however, is usage. I think maybe once in several months did I even get close to hitting my limit with the $20 plan.
With Claude, however, I feel like I can sneeze and half my weekly usage is gone. Same $20 price tag.
That's been my experience, I'm sure it differs user to user though.
nice, I have been hearing about Codex a lot, so I will def try it
I am using both claude code and github copilot at same time.I like using copilot actually
Perhaps you could try GLM5.1, I see the offical demo, it's nice
I've gone fully custom and would never look back. All of the popular tools lack the experience I'm after, generally lacking the advanced customization I have found makes weaker models more capable. I'm also moving towards open weight models.
The underlying motivation is to avoid Big Ai as much as possible. Incentives are not aligned.
> the advanced customization I have found makes weaker models more capable.
Like what?
Good prompts and tools. When I first went down this path, I grabbed the Claude Code prompts and adapted them. Doing so made gemini-3-flash as good as pro for 80%+ of my normal tasks. I became spoiled with the turnaround time (much faster) and now become impatient with the mega models.
I can also do things like out the AGENTS.md contents in the system prompt and change file read to put contents there as well. Thai avoids duplicates in the messages when they fail at editing and decide to reread the file
Just pay the $20 for codex and use it. It's the only real alternative.
opencode, kimi-cli, qwen-cli, pi-agent,