The AI map of 2026: who's who

Open any AI leaderboard and you start to drown. Dozens of names, a number on the end of each one, every one of them promising to be the best. It feels like walking into an electronics shop with forty identical boxes on the shelf and no assistant in sight.

Behind that noise sits a fairly simple map. Ten minutes from now you'll be carrying it around in your head, and leaderboards will stop being frightening.

Four words so the news stops confusing you

Half the confusion in AI articles comes from four different things sharing one name. Let's separate them.

• Model — the "brain" that does the thinking. One company has several: faster and cheaper, slower and smarter.
• Chatbot — the window you talk to the model through. ChatGPT, Claude and Gemini are windows.
• Version — the generation of a model. The number on the end grows roughly every six to twelve months, and you don't need to chase it.
• Subscription — paid access: the senior models and lifted limits. On a free tier you get the junior models and a ceiling on how much you can ask.

These words come up dozens of times later in the course, so it's worth coming back here if they blur together.

The three families that hold almost everything

The easiest way to think about them is as three acquaintances you go to with different favours. Each one can do nearly everything, but each has its own character — and that character isn't an accident, it grew out of the company's history.

• ChatGPT, from a company called OpenAI — the best known and the most all-round. Text, voice, images, reading your files, its own store of ready-made bots. It has the most users, which means the most guides, videos and tips online: whatever you're about to ask, someone has asked it before you.
• Claude, from Anthropic — the one people praise for language. It's more careful with text, calmer in tone, doesn't scatter exclamation marks, and holds a long document in its head better. Programmers love it for its own reasons.
• Gemini, from Google — it lives where you already spend your day: mail, documents, spreadsheets, search. If you have Gmail, you already have Gemini; there's nothing separate to sign up for.

Any one of the three is enough for an ordinary working day — emails, writing, reading documents, ideas. The gap between them is smaller than the gap between "I use AI" and "I don't". We'll get to the shades of strength in the next two lessons.

Where their characters come from

The characters aren't accidental; they grew out of the companies' histories, and in two lines it goes like this. OpenAI was a quiet research lab from 2015 until it published ChatGPT on 30 November 2022 and everyone started talking about artificial intelligence. Anthropic was founded in 2021 by researchers who left OpenAI believing safety mattered more than shipping speed — hence Claude's caution. Google took the longest road and went through science: its DeepMind lab beat Lee Sedol, one of the strongest Go players alive, back in 2016.

The specialists: who you call when the all-rounder gets stuck

Around the big three lives a crowd of narrower models. They don't try to do everything at once; they do one job, and do it in a way the all-rounders can't match yet. You don't need to sign up for any of them today, but it helps to know they exist: sooner or later you'll see a friend's result and want the same thing.

• Midjourney — images. It began as an experiment inside Discord, a messenger popular with gamers and developers: people typed a phrase, a bot posted four pictures, and the whole crowd watched each other's results. Out of that grew one of the most recognisable image models around.
• Perplexity — search that answers in words and shows you where it got them. Not ten blue links, but a short answer with footnotes you can click and check.
• Suno — music. You write a line about the song you want and get back a finished track with vocals. First encounters usually end with someone spending half an hour making songs about their cat.
• DeepSeek — an open model from China. Open here means something narrow: the company published the weights — what came out of training — and you can download and run them on your own machine without handing anyone a single line of your data. The training code and data are almost never published, so «open» does not mean you can see how it was made. For companies that's sometimes the argument that decides everything.

Short version: the all-rounder is your main tool, a specialist is something you go to deliberately, for one specific job.

Why "which one is best" has no answer

Asking which AI is best is a bit like asking which tool is best when someone is about to fix a tap. A screwdriver? A wrench? Depends what's leaking.

The useful question sounds different: best at what? Write an email, work through a hundred-page contract, find fresh statistics with sources, draw a logo — four jobs, four different answers. People who use AI every day usually keep a combo: one paid all-rounder as the main one, plus the free tiers of the others for the things they're good at. Nobody picks one for life.

The good news: you can hardly get this wrong

What you're really learning isn't the buttons of one particular service — it's how to explain a task in words. That skill is called prompting, and it carries between models without losses: learn to phrase things for ChatGPT and tomorrow you'll talk to Claude the same way.

So choosing the "wrong" model costs you nothing. None of the time you invest goes missing.

How to read leaderboards without losing your mind

Sooner or later you'll come across a table with models ranked in order, and you'll want to just take the top row. It's worth knowing what such a table measures.

The best known of the public rankings is LMArena, and it is built more honestly than most. The scheme is simple and rather beautiful: a person is shown two answers to their own question, with no names attached, and picks the one they liked more. Only after the vote do the names appear. Thousands of blind comparisons like that add up into a table — so what lives in it isn't company promises but the actual preferences of real people.

Even so, it has weak spots. First, people vote for the answer they liked, and what people like is long and nicely formatted — not necessarily correct. Second, companies run trial versions of their models through the arena and show the best one. And above all, the table shows an average over other people's tasks. The model that wins the arena may well lose to its neighbour on your emails and your documents, because nobody was voting on your work. So leaderboards are good for curiosity, and the decision gets made on your own scenarios. How to work those out is the next two lessons.

The myths that get in the way of choosing

These three come up in every conversation about AI, and all three are wrong.

• "There's one best model and the rest are copies." The three have different creators, different data and different principles — they overtake each other in different disciplines, and the picture changes several times a year.
• "A paid tier from one is always better than a free tier from another." Not necessarily. The free versions of the big three today do more than the paid versions did two years ago.
• "By the time I choose, it'll all be out of date." The models do get updated. The skill of talking to them doesn't go out of date at all.

Which language to talk to them in

Your own. All three read and write fluently in dozens of languages, idioms and heavy official phrasing included — the kind you'll later ask them to strip out. Write in the language you think in: the quality gap between languages is smaller than the gap between a clearly framed task and a vague one.