Language Models (LLM)

Where to start, if you are starting today

The question asked most often here is "which model should I pick", and it is almost always premature. The differences between the three leading models are far smaller than the marketing implies, and they are smallest of all for someone who does not yet know what they want to do. A beginner gets more out of picking one and using it for a month than out of comparing three for a week.

A practical order: open one of them and give it a real task you were going to do anyway — draft an awkward email, summarise a contract, explain an unfamiliar function — then compare the result against what you would have produced yourself. If it saved you time, carry on. Comparisons and pricing matter at the next stage, once it is clear what you are running and at what volume.

Three questions that decide the choice

• What is your volume. Daily personal use fits inside the flat monthly subscription almost every time. A system making thousands of calls runs against the API, where the pricing model and the arithmetic are completely different.
• Is your content sensitive. If documents are not allowed to leave the organisation, the question stops being which model is cleverest and becomes what you are permitted to run at all. That is where local models come in — an entire category most people do not know exists.
• Are you using or building. A user needs a good interface and good prompts. Someone shipping a product needs an API, a view of cost per call, and a way to measure whether the output is actually correct. Three quite different subjects.

What to know before going further

Mistakes that keep repeating

Asking a short question and expecting a good answer. The gap between a mediocre result and an excellent one is decided almost entirely by how much context you supplied, not by which model you used. "Write a client email" against "write to a client who has cancelled twice, friendly but direct, offering two specific times" — same model, different world.

Assuming it remembers. A new conversation starts from nothing. What you explained yesterday is not there, unless the tool keeps memory explicitly — and even then, only partly.

Asking it to do arithmetic. A language model completes text; it is not a calculator. Totals, percentages and date differences need checking, including when the answer looks entirely reasonable.

Giving up after one bad attempt. The first prompt is almost never the right one. Two rounds of correction are the normal case, not a sign that something is broken.

What each cluster above is for

The four groups are not just a tidy split — they are ordered by the stage you are at, and they are worth reading in that order rather than jumping:

• A guide to each model — the first stage. Each guide covers what that model does well, where it weakens, and what daily use of it actually looks like. If you only want to get going, read one and stop; the rest makes far more sense once you have hands-on experience.
• Compare, price and choose — the stage you reach once use is regular and you start wondering whether you are overpaying. This is also where cost-by-volume is worked out, and it is the part that surprises most people moving from personal use to a running system.
• Local and free — a category that serves two very different people: the one who does not want to pay, and the one who is not allowed to send their content anywhere. The second is the real reason the category exists.
• Prompting and building — for writers and developers. If you are not shipping a product, the relevant half here is prompt writing; the API can wait.

And if you are unsure which cluster you belong to: most people are in the first two and imagine they need the third and fourth. There is no point installing a local model before it is clear that the hosted one does not meet the need.

What you will not find here

There is no "best model" ranking on this page, and no benchmark table copied out of a press release. Benchmark scores shift with every update, are measured on tasks chosen by the people publishing them, and predict less than you would think about whether a tool will work for you. What is here instead: a guide per model, a comparison that explains where they genuinely differ, and pricing you can actually calculate with.