An LLM does not know things the way a person does. It calculates what is likely to come next, at a scale that changes everything: by seeing billions of sentences, the model learns not only language but what the sentences are about. What a quote looks like, what a notice period is, how to answer a complaint politely. ChatGPT runs on OpenAI's models, Claude on Anthropic's, Gemini on Google's, Le Chat on those of the French company Mistral.
The breakthrough came in 2017 with Google's paper "Attention Is All You Need", which described the transformer: a way for a model to weigh every word of a text at once. OpenAI built its GPT models on it, and on 30 November 2022 released ChatGPT. For the first time anyone could talk to a language model.
How the model picks the next word
Text is cut into tokens. In English a token is roughly three quarters of a word; other languages cost more, Dutch for instance averaging 1.67 tokens per word against 1.23 for English, so the same email costs a little more in some languages than others. For every possible next token the model calculates a probability, picks one and starts again. What you give it, a role, an example, your own documents, decides which continuation it finds most likely. How much text it can read at once is its context window.
The figure shows why that matters: without your real price list, the model picks a figure that sounds right. That is an AI hallucination.
A line from a quote, and what the model considers next
Our price for the complete bathroom comes to
- €42%
- about21%
- in total12%
- excluding9%
- after4%
Illustration. Without your real price list in the context, the model picks an amount that sounds likely, not the one that is correct.
What a thousand emails cost
The models have become cheap. You pay per million tokens, separately for what goes in and what comes out.
| Model | Price per million tokens (in / out) | 1,000 emails a month |
|---|---|---|
| Claude Haiku 5.5 or GPT-6 Luna | $0.10 / $0.50 | about $0.34 |
| Gemini 3.8 Flash | $0.75 / $3.75 | about $2.51 |
| Claude Sonnet 5.5 or GPT-6.1 Sol | $2 / $10 | about $6.70 |
| Claude Opus 5.5 | $4 / $20 | about $13.40 |
The model is rarely the big cost. Building is: connecting it to your inbox or software, giving it your own documents, and making sure it does not invent things. Estimate your own usage with the AI cost calculator; what a complete assistant costs is in AI chatbot cost, and how tokens are counted under AI tokens.
Where a language model saves time, and where it does not
An LLM is a fast first draft, not a final check.
- Writing
- Strong atA first draft of an email, quote text or post in your tone
- Weak atFacts it was not given: prices, dates, names
- Reading
- Strong atSummarising long emails, contracts or reports
- Weak atCopying figures exactly without a check
- Sorting
- Strong atRouting enquiries, pulling data out of text
- Weak atDeciding what is legally or financially correct
- Up to date
- Strong atWorking with documents you supply
- Weak atKnowing what changed since its training
Give it your own sources, and have a person read what goes out.
Where to start
Pick a frequent task where a mistake is not a disaster: sorting email, drafting a quote, summarising reports. Put a ChatGPT, Claude or Gemini business plan against it and measure for a month how much time it saves; ChatGPT vs Claude compares the two most used. If the model needs to know your own information, give it that with each question through retrieval-augmented generation rather than retraining it; the difference is under fine-tuning. In the EU, the AI Act has required AI literacy among staff who use these tools since February 2025; the EU AI Act guide sets out what that means.
What owners ask about language models
What is the difference between an LLM and ChatGPT?
The LLM is the model, the engine. ChatGPT is the application you use to talk to OpenAI's models, with extras such as web search and reading files.
Does an LLM learn from my data?
On business plans and the API, the major vendors say they do not train on your data. Read the terms of your plan, and keep sensitive data out of free accounts.
Which model should I choose?
It depends on the task. A small, cheap model handles sorting; harder writing and reasoning need a larger one. Test two or three on your own examples.