Sonnet, Opus, Sol or Luna: which model should you pick?

Almost nobody picks a model. You open Claude or ChatGPT, one is already selected, and that is the one you use. A shame, because the difference between those models is bigger than the difference between the brands. And if you pay per use, the price difference runs to over forty times.

I turned it into a short questionnaire. Nine questions, and you get one model with the price, the strengths and the place where it lets you down. First some explanation, because the names tell you nothing on their own.

First question: where are you choosing?

This matters more than it looks, which is why it is the first question in the list.

If you pick your model from the dropdown in Claude or ChatGPT, you see a handful of options. You pay a fixed amount per month, usually around €20, and your usage makes no difference to the bill. The choice is then purely about quality and speed.

If you are building something, or having something built, it is an entirely different story. Then you pay per use, and suddenly there are cheap models available that you never see in the app at all. That is where model choice becomes a budget line.

That difference is why people talk past each other on this subject. They are not discussing the same problem.

What does a model actually cost?

Prices are quoted per million tokens. A token is roughly a piece of a word. A million tokens comes to about 750,000 words, or a stack of paper of more than 1,500 pages.

You pay one rate for the input and a higher rate for what the model writes back, usually a factor of five or six. Which makes sense when you think about it: writing takes more computation than reading.

To put numbers on it. Have 1,500 pages of text written and the cheapest model costs about $1 while the most expensive costs $50. For one person doing the odd thing with it, that is negligible. For a system running thousands of times a day it is the difference between €50 and €2,500 a month.

The seven this is about

There are two families, and within each family it climbs from fast and cheap to slow and expensive.

At Anthropic you have Haiku 4.5 as the fast, cheap one, Sonnet 5 as the model most people should be using most days, Opus 5 for work where mistakes cost money, and Fable 5 right at the top for jobs that have to run on their own for hours.

At OpenAI that is Luna as the fast, cheap one, Terra as the middle model and Sol as the flagship. Of all seven, Sol is the strongest at actually carrying things out: browsing, driving systems, filling in forms.

The middle models of both families, Sonnet 5 and Terra, cost practically the same and perform about the same. So you choose there on where you already are rather than on benchmarks.

What they really differ on

How clever. The obvious one, and also the least useful measure. Most of the tasks people do are not difficult at all. Summarising an email goes as well on the cheapest model as on the most expensive.

How fast. If somebody is waiting for the answer, speed is not a luxury. The fastest models answer almost immediately; the slowest thinks for a while before it starts writing.

How much text goes in at once. Most models swallow a million tokens, so around 1,500 pages, in one go. Two exceptions matter. Haiku 4.5 stops at 200,000 tokens, so around 300 pages. And Luna claims a million but retrieving information from a slab of text that size goes badly wrong. For document analysis that is the wrong tool, however tempting the price.

The number everybody skips

Every model has a date after which it knows nothing. That sounds like a footnote, but in practice it is the number that gets in your way most often.

The spread is wide. Opus 5 runs to May of this year. Most others stop in January. Haiku 4.5 knows nothing after February 2025, which is eighteen months ago.

If you work with legislation that has just changed, or with software that moves quickly, the oldest model will invent answers with complete conviction that are no longer correct. It does not tell you it is guessing. Supply the current information yourself and none of it matters.

How to read the outcome

You get one model with the price, the specifications and a block on where it lets you down. Here too, each of these seven is not good at something, and you want to know that before you put anything into production.

Below it there is a second model. That is not the runner-up but your counterpart. If an expensive model came out, that line tells you which cheaper model should be handling the bulk of your work. If a cheap model came out, you see where to escalate when something gets stuck.

That is the best advice I can give on this subject, incidentally. Do not put one model on everything. The saving is not a cheaper flagship, it is not pushing work through the flagship that a cheaper model finishes identically.

Right at the bottom there is a price ladder. All seven on one line, Anthropic above and OpenAI below, ranked by what they cost to have them write. That shows you at a glance how far apart they are and how close some of them sit.

Small print

Prices, context windows and knowledge cutoffs are as of 23 August 2026 and come from the documentation of Anthropic and OpenAI themselves. These rates move fast. OpenAI cut the price of two of its three models sharply at the end of July, and on 21 August temporarily cut the flagship as well. Always check the current rates before you put anything into production.

I left one model out. Claude Mythos 5 is available by invitation only to a small number of organisations, so as a choice it does not exist for most people.