Choose a model the way you staff a job.

Open the model picker on any serious router today and you are staring at four hundred options. OpenRouter puts the number on its homepage: 400+ models, across seventy-odd providers, behind one API. It is presented as abundance. It reads as an exam.

The moment you have four hundred choices, someone expects you to compare them. Context window against context window, benchmark score against benchmark score, tokens per second for the ones where you have started to care about tokens per second. You did not sign up to run procurement for a language model, and yet here you are, squinting at a leaderboard, working out whether the model two rows down is worth the latency for a task you have not fully described even to yourself.

This is the first wrong answer, and it is the industry default: make the director read the spec sheet. It turns delegation back into operations. You wanted to hand off the work. Instead you are interviewing your staff on their benchmark scores before you have told them what the job is.

The scores answer a question you are not asking

Think about how you would actually assign this to a person.

One associate drafts replies to a sensitive client, the kind where a wrong sentence costs you the account. Another skims a news feed and pulls the three things you should know before lunch. You would not put those two through the same interview, and you would certainly not ask either for their score on a reasoning benchmark. You would ask what it costs to be wrong, and you would staff accordingly.

The client replies get your most careful hand, checked before anything leaves the building. The news skim gets something fast and cheap, because the worst case is a summary you glance past. Stakes decided that. Raw capability barely entered it. Benchmarks rank models on average difficulty across tasks that are not yours. Stakes rank them on the one axis you actually own: what happens when this associate gets it wrong.

The second answer is worse

Faced with a director frozen in front of four hundred rows, the industry has a fix ready. Let us pick for you. Auto-routing: the best model for every prompt, chosen behind the glass, no dropdown required.

It is a relief, and it is a trap. The confusion was real. The cure removes the wrong thing. It does not remove the spec sheet. It removes you. The decision still gets made on every single call. You simply no longer know what it was, and cannot tell whether the sensitive client reply went to your careful model or to whichever cheap one happened to be winning the router’s math that second.

A director who cannot see who they staffed a job with is not directing. That is the entire role, gone quiet. Hiding the choice leaves you a spectator to your own operation, which is a worse seat than operator ever was.

Make the call in words, and keep it in view

There is a version that respects both facts at once: the choice is real, and you should not need a degree to make it.

You say, in plain language, what an associate is for and how much it would cost to be wrong. The careful one handles anything a client will read. The fast one handles the digests nobody will ever sue you over. You can see at a glance which model each associate is running, and you change it when the work changes, because the work always changes. No leaderboard, no JSON. Just a sentence.

This is the bet we are making. The model is a rental, and that is fine; you rent the engine and keep the judgment about which engine goes where. That judgment is a management skill you already have from working with people, and it outlasts every model release and every new name at the top of the board. When the frontier moves next month, and it will, you re-staff in a sentence. You do not go back to school.

Four hundred models is not a menu to be read. It is a labour pool to be directed. The tools that made you study it and the tools that studied it for you in the dark got the same thing wrong. They forgot you already know how to assign work. You were just never allowed to.