Skip to content
Blog
Not every question needs the expensive model

Not every question needs the expensive model

Oct 8, 2026
Part 11 of the Journey

In the previous post I wrote about the token-saving rule: the agent answers with the result only. No “ok”, no “done”, no courtesy line. This part is the other side of the same bill. The thing that picks the path also does not need to be the model that writes a paragraph.

What I wanted to solve

I work during the day, and at night and on Sundays I build with AI agents. In Hermes, the agent program I use, each agent has a role. Tiny questions show up all the time. Who handles this? Is this a yes or a no? Go or stop?

If I send every one of those to a large chat model, it tends to explain, wander, and spend tokens. A token, in one sentence, is a small piece of text the model reads or writes. Billing and quotas move by those pieces.

I did not need an essay. I needed a decision.

The rule I wrote was direct: simple decisions go to Jev or to Clef.

I use Jev. Clef, at the date of this part, was still in testing. The Jev cost figure below comes from my own cost study on 24 September 2026. It is not a published independent test. Read it with that discount.

Quick summary

  • A closed question (yes or no, who does the task) does not need the model that generates long text.
  • Jev, from TypeSafe, is a fast judge linked to the Hermes agents to pick paths.
  • In the 24 September 2026 cost study, Jev came out at $0.042 per million input tokens. Output is free.
  • Clef, Cloudflare’s decision model, uses the same question format. On 3 October 2026 it was still in testing.
  • Split “deciding” from “thinking”.

First: what routing is

Routing, in this post, is only this: look at the request and pick the path.

Picture a building front desk. The person at the door does not draft the contract. They ask whether you came for the second floor or the third, then point down the hall.

The real mechanism is drier. A program takes the text, matches it against a fixed list of options, and returns one of them. The analogy stops here: a human receptionist still talks. Jev and Clef do not talk, do not explain, and do not build a paragraph. They return the option.

That only works if the question is closed. A closed question is one where the possible answers already exist before you ask. Yes or no. Which agent. Which queue.

A daily example: sorting mail on a table. A bill, a note from a friend, junk. You do not write a letter to classify each envelope. You look at what is in front of you and put it on the right pile.

A large chat model can mark the pile too. It was trained to write, though. It often thinks out loud, bills you for the output, and uses up quota. For a repeated decision, that costs money and does not help.

If the answer already fits in a yes, a no, or a name, paying for a paragraph does not make sense.

What I used

I used Hermes, with the agents already in this series. I used Jev for simple decisions. I tested Clef with the same question format. And I used a written rule: a simple decision does not go to the expensive model.

Logins, plans, and accounts stay out of this text. For you, the rule is enough.

Jev is TypeSafe’s System One model. In their naming, System One points to a fast judge: it reads the text, picks among ready-made options, and stops.

In the Hermes agents, that is exactly where it sits: pick paths and answer closed questions. It does not write the post. It does not approve in my place.

Think of an answer sheet, not an essay. You send the prompt and the choices. You get the marked option back. If the question needs a paragraph, you are holding the wrong tool.

Clef is Cloudflare’s decision model, on Workers AI. Workers AI is Cloudflare’s service for running AI models on its network.

The question format is the same as Jev’s: a note, closed options, an answer a program can use without fishing a phrase out of a blob of text.

On 3 October 2026, Clef was in testing. The first test worked. The test script, the little program that sends the question and reads the answer, still needed work. So I do not treat Clef as a finished piece.

I do not have a confirmed cost number for Clef. No checked number, no figure in this post.

How it went

This was not an install tutorial. It was a design choice, after looking at cost.

If you want to copy the logic, the path is this:

  1. Write down the questions your agents ask all the time. “Who handles this?” and “is this a yes or a no?” show up a lot.
  2. Mark which ones already have a closed list of answers. Those are candidates for the cheap judge.
  3. Leave the large model for work that needs reasoning: writing, explaining, judging a new case.
  4. Link a fast judge (in my case, Jev on the Hermes agents) to the closed questions.
  5. If you test a second judge in the same format, as I did with Clef on 3 October, treat it as a test until the test program is stable. The first test worked; the tweak was not done yet.

When this becomes routine, the writing model only comes in for work that actually needs text.

What went wrong

Clef was not solid yet. First test ok, script still being adjusted. Calling that “done” would be a lie.

The other lesson is habit. It is easy to send everything to the model already in the chat. It answers. Answering and answering at the right price are not the same thing.

The fast judge also does not guess what you failed to write. If the option list is crooked, the mark comes back crooked, sometimes looking sure of itself. The list is still my job. None of these tools replace judgment. They shorten the path between “I know what I want” and “it is live”.

Costs and limits

In the cost study of my set of tools, on 24 September 2026, Jev was $0.042 per million input tokens. Output is free. Unit: US dollars per million tokens read. Comparison base: that study, on that date. Source: the study itself. I will not invent the large model’s price on the other side to make the contrast look prettier.

What Jev does well: repeated decisions, closed questions, picking an agent path.

What it does not do: write the post, explain a topic to a reader, approve in my place.

Clef, in this part of the story, was in testing. No confirmed cost.

Cheap routing only saves quota if you actually divert the small decisions. If the closed question still goes to the expensive model, the rule on paper does not cut the bill.

What I would do differently

I would start with the list. Before wiring any judge, I would write the closed questions and the allowed options. Without that, the cheap model has nothing to mark, and the temptation is to throw the doubt back at the model that writes.

I also would not call a second judge ready just because the first test passed. With Clef, the first test worked and the script still needed work. I would keep the “in testing” label until that part was done.

Who it is for, and who it is not for

It is for you if you already have more than one agent, or if you notice the expensive model being called just to say yes or no.

It is not for you if you use AI once in a while, in a single chat, still in the “ask a question” mode. The gain is small, and the extra part gets in the way.

It is also not for deciding, on its own, anything that touches personal data or an approval that is mine. I check.

Closing

I wanted to stop spending expensive reasoning on front-desk questions. Jev already does that job for the Hermes agents. Clef, in testing, tries the same format on Cloudflare. The rule to copy: if the answer fits in a yes, a no, or an agent’s name, do not call the model that writes paragraphs.

In the next post I will cover the site address and the choices: domain, Hugo, GitHub, and Cloudflare.

If you want to follow the journey, I am on X: @RodrigoesPorto.

Sources and links

Token Economy: Why My Agents Only Reply With ResultsThe site address: domain, Hugo, GitHub and Cloudflare