Part 2: Choosing a Language Model — Quality, Cost, and Speed

In part one we took apart the five pieces that make up an assistant, and the thinking engine was one of the three that must be there. Now it is time to pick that engine.

There are plenty of language models around, and almost all of them look clever in a demo. The trouble is that cheap ones are usually less careful, the smartest ones are sometimes slow, and the fast ones sometimes answer without checking. You rarely get all three at once.

Quality is not about a ranking, it is about being good enough

Scores from various exams make interesting reading, but they do little for you. Expensive models win on hard work: reading long documents, laying out ordered steps, or writing code. For light work — tidying notes, writing summaries, replying to routine messages — a mid-range model is usually enough, and the result is not far off.

A more useful test: gather five to ten examples of real work, run them on two models, then compare the results. What matters is not which model wins overall, but which one is wrong least often on your work.

Cost has two shapes, and they behave differently

Some providers charge by the number of words going in and coming out. Others use a flat monthly subscription, no matter how much you use. If your assistant runs every day and handles many tasks, a subscription is usually easier to predict at the end of the month.

Worth remembering: big spending usually comes from habits, not from the model itself. Work that repeats every day is better turned into a scheduled job, so it is not called again and again. A small model can prepare the raw material, and a more careful model can finish it.

Speed shows up when you are waiting

When you chat through WhatsApp or Telegram, a pause of a few seconds already feels long. For conversation, a fast model is more comfortable — the answer lands right away, even if it is sometimes shallow. A model that thinks for a long time suits work nobody waits for: drafting the evening report, summarising an archive, or checking files for tomorrow's meeting.

So there is no need to lean on one model for everything. Use the quick one to talk, and the careful one for heavy work.

How to choose without going in circles

Write down the three jobs your assistant will do most often.
Start with one mid-range model and use it for a week or two.
Note when the result is off — that tells you when to move up a class.

Those notes are more useful than a price list. If it turns out nothing went wrong, you do not need to pay more yet.

One last thing: prices and speed change, but your needs stay fairly steady. Choose based on the work you actually do, not on whatever came out most recently.

Next part: where to run the agent — small servers, monthly cost, and why an assistant has to stay on.