Let's extend the codex rate limit ourself with a little trick

Many codex users don’t know when to use which model and that leads to way higher usage than neccessary.

Let’s try something simple that should help.

OpenAI has a new model - a so called decission model. Let’s use a MCP server that decides which model to use.

It works like this:

You install the mcp server (tell codex to do so) and then tell it to create a skill that says “always use the mcp server to identify which model to use for the given task then start a subagent with that model to work on that task”

I think we can run codex on the cheapest model for that and the mcp server will answer e.g. with “use ASTRA ultra for that” or “use sol 6.1 on medium” .. and so on…

I have asked codex to build that model router - anyone willing to give it a try? It is open source!

@jochenschultz Can you explain where those numbers come from? As I understand it, they are the main source when deciding which model to choose.

"max_complexity": 0.55,
"capabilities": {
 "general": 0.62,
 "documentation": 0.85,
 "testing": 0.68,
 "frontend": 0.68,
 "implementation": 0.65,
 "refactoring": 0.59,
 "debugging": 0.57,
 "concurrency": 0.4,
 "security": 0.42,
 "architecture": 0.42
}

I am also not sure that this division of tasks is right. For example implementation, Depending on what you implement, you may need a different model, like doing CRUD requires a simple model, while designing a new class may require a more complex model.

you might have to adjust that and extend the mcp so you can give feedback and let them be adjusted. it is not a finished project .. or do you see a price tag?