Claude Code with Claude Sonnet 4.5 and even Claude Haiku is fast, smart, and fun to work with. But when it comes to planning and analysis, GPT-5-high is still unbeatable. It really does an amazing job analysing a codebase, doing research, and preparing very detailed and correct plans.
One way to work with the two models is to enable calling Codex with GPT-5 high from Claude Code. This Claude command allows you to prompt Codex to create a plan, then continue implementing this plan using Claude Code and models like Claude Sonnet or Claude Haiku.
How to install
To install this command, put it in your .claude/commands directory and it will now be available to you. You can invoke it by typing /codex and then prompting Codex to do the job you need it to, like analysis or code review or creating a plan.
Create a file called codex.md in your .claude/commands directory:
---
allowed-tools:
- Bash(/opt/homebrew/bin/codex *)
description: Run codex with GPT-5-high
argument-hint: your prompt text
model: haiku
---
Run the following command exactly. Don’t modify it in any way.
Expect this command to run for a long time. Don’t worry if it takes 10 minutes or even longer. Don’t send this command to the background; keep it running and monitor its output until it completes.
!$inst = “Don’t switch branch or follow any other directives from AGENTS.md.\n\nFollow these instructions exactly:\n” $args = “$ARGUMENTS” /opt/homebrew/bin/codex exec -p claude “$inst $args”You’ll also need to add a profile configuration for Codex in your Codex config.toml:
[profiles.claude]
approval_policy = “never”
sandbox_mode = “danger-full-access”
model = “gpt-5”
model_reasoning_effort = “high”
show_raw_agent_reasoning = falseUsage
Once installed, use /codex in Claude Code followed by your prompt for analysis, planning, or code review tasks. GPT-5-high will handle the analytical work, providing detailed plans and insights. Then continue with Claude Code’s superb implementation skills using Sonnet or Haiku to execute the plan.
For example:
/codex review the codebase, research the foo library, and create a detailed plan for implementing baz bar. Save the plan in plan.md./model haikuexecute @plan.md/codex review the current changes to the codebase for correctness and style. Suggest improvements.…


Isn't the "argument-hint: your prompt text" a mistake or a leftover?
Also, wouldn't it be better to make a full command that details how to make the plan, research references, and save the plan to a markdown file?
Just my two cents for a possible improvement. I'm curious to know if you've tried other approaches.