A former Cursor power user who switched to Claude Code has now moved to Codex as his daily driver. While the agents are converging in capability, Codex wins on pricing efficiency (GPT-5 costs a third of Sonnet and a tenth of Opus), more generous usage limits, and superior GitHub integration with auto code review and inline bug fixes.
The AI coding agent space is converging rapidly. Cursor set the foundation, Claude Code made improvements (like to-do lists and better diffing), Cursor copied those, and then Codex copied them too. The author notes that Codex is so similar to Claude Code that he wonders if they trained on Claude Code's outputs.
Key differences in how the agents work:
"I kind of like knowing the same company that made the tool I'm using also trained the models. So they know how to use it best."
Both services offer free, ~$20, and $100-200 tiers. However, the underlying economics favor Codex because GPT-5 is significantly more efficient than Claude's models:
"The number one complaint I see with coding agents is running out of credits. So I think Codex has a very strong edge here."
Beyond coding, ChatGPT subscriptions also include image generation, video generation, and a more polished desktop app compared to Claude's "slow Electron web wrapper."
Claude Code and Codex have similar terminal UIs (Claude's is slightly better), but Claude's permission system is a major pain point. The author always launches Claude Code with "dangerously skipped permissions" because settings don't persist between sessions.
Codex, by contrast, recognizes when it's in a git-tracked repository and is permissive by default.
Claude Code offers more features: sub-agents, hooks, extensive configuration. But the author learned from using Claude Code that he doesn't actually care about features:
"I just want the best agent that I can ask for what I want and get it most reliably. All I find I need is an agent and a good instructions file."
One frustration: Claude Code only supports claude.md for instructions, while Cursor, Codex, and Builder.io all support the agents.md standard.
The GitHub integration is what puts Codex over the top. Claude Code's GitHub app was disappointing - verbose reviews, didn't find obvious bugs, couldn't fix issues via comments.
Codex's GitHub integration:
The consistency matters: when you build expertise with Codex in the terminal, that knowledge transfers directly to the GitHub integration.
At Builder.io, they've integrated Codex to enable designers and PMs to make updates directly:
"Our designers just jump into Builder.io and use Codex to make updates to our sites and apps through prompting and through our Figma-like visual editor."
The whole team - designers, PMs, developers - builds off the same foundation with the same codebase, same models, and same agents.md instructions.
"I was a Cursor agent power user for months... but then Claude Code came out and that became my go-to. But already my go-to has changed again."
"I learned from Claude Code that I really don't care about features at all. I just want the best agent that I can ask for what I want and get it most reliably."
"In 2025, I really don't like to deal with handoffs anymore."
| Aspect | Codex | Claude Code |
|---|---|---|
| Pricing efficiency | Better (GPT-5 costs less) | Higher cost per token |
| Usage limits | Very generous | Users hit limits often |
| Features | Basic | More (sub-agents, hooks) |
| GitHub integration | Excellent | Disappointing |
| Permission system | Permissive by default in git repos | Requires manual override each session |
| Instructions file | agents.md (standard) | claude.md only |
| Terminal UI | Good | Slightly better |
| MCP integration | Basic | Better with click-to-install connectors |