I strongly recommend Claude Code over Codex for almost everybody. The big issue for me is speed.
With Codex, I average around 30 tokens per second on the flagship models I use. With Claude Code, I average around 150 tokens per second.
What does this mean in practice? Well, as a power user who’s usually running up to 15 agents in parallel, 15 Codex agents at 30 tokens per second add up to 450 tokens per second. Three Claude Code agents at 150 tokens per second add up to the same total.
Those are my reported rates, not a controlled benchmark. The arithmetic is about aggregate token generation, not equal code quality or five times as many completed tasks. Models and settings change rates, and faster modes can use more of your allowance or cost more: https://developers.openai.com/codex/speed/
Running 15 agents in parallel is demanding #
People think, oh, the agent is doing the work. It doesn’t matter how long it takes. But the reality is, it does. Running 15 agents in parallel is demanding. It’s not easy work. My productivity takes a significant hit when I’m coordinating that many agents.
When the token-rate arithmetic is 15 agents with Codex versus three with Claude Code, you know which one I’m going to prefer. Benchmark scores and project results don’t show how demanding that coordination feels.
You just feel like, oh shit, Claude Code just gets shit done. For me, it’s not about the system prompt. It’s not about the skills. It’s not about the tools. It’s the speed.