September 14, 2026 · 5 min

What a narrow model actually saves you

The case for small models used to be easy: frontier models were expensive, so anything narrow and cheap won on price. Then the frontier got cheap. OpenAI cut Luna’s output price by 80% in July, and the gap that made the argument obvious mostly closed.

So we did the arithmetic honestly, against the prices as they stand today, and it turns out the money argument is weaker than we would like — and the argument that replaces it is better.

The prices

Published list prices, September 2026. Batch and cache discounts apply to the frontier rows and would narrow the gap further.

What that is worth in practice

Take the job our first model does: rewriting a thousand-word article so it reads like a person wrote it. That is about 1,300 tokens in and 1,150 out.

• On meraGPT: $0.00024
• On Luna: $0.0017
• On Haiku 4.5: $0.0071

Seven times cheaper than the cheapest frontier option, twenty-nine times cheaper than the most expensive one here. Which sounds decisive until you notice the units. A hundred articles a month is two cents against sixteen. Nobody switches providers for fourteen cents.

The money only becomes real at volume or in a loop. Ten thousand articles a month is $2.37 against $16.40 against $70.50. A pipeline that restyles every page of every customer’s site on a weekly schedule is the shape where that compounds into something you notice on an invoice. If you are not in that shape, be honest with yourself: the price is not your reason.

The reason that survives

Here is the part that does not depend on volume. On the job itself, the narrow model is better.

We measured all three frontier models on restyling, scored so that 0.50 means indistinguishable from human writing and 1.00 means obviously machine-written. Haiku 4.5, Flash Lite and Luna all landed around 0.90. Our model brings the finished text to 0.65, including on output from a model it was never trained on. The full comparison is here.

That is the actual trade, and it runs the opposite way to the one people expect. You are not paying less to get less. You are paying less because the model does not also know how to write Rust, argue about Kant, or plan a wedding — and on this one task, the thing that only does this task wins.

The same held when we built the second model. A frontier model prompted to predict the searches an assistant would run scored 0.082 lower on coverage than ours. We will be straight about the caveat: it returned two queries where ours returned six, and the metric rewards more attempts, so some of that gap is budget rather than skill.

When a frontier model is the right answer

Often. If you are doing something once, or exploring, or the task changes every week, a general model is the correct tool and the fourteen cents is irrelevant. Narrow models are worth it when the task is fixed, runs constantly, and someone can tell the difference in the output.

That is three conditions, not one, and plenty of work fails at least one of them. We would rather say so than sell you a model you will stop using in a month.

If your work does meet them, you can try both models without an account, or read the API docs. Pricing is on the pricing section, and it is the same $0.05 / $0.15 for every model we ship.