What is a large language model? A plain-language guide
A large language model is a program that has learned patterns in text well enough to continue, answer, summarise and translate it. This guide explains how they are built, using the research papers that defined the field and numbers from our own price and context tables. For the numbers behind the bills see the AI costs section; to see how text becomes tokens, use the token counter. Section hub: AI models explained.
Prices last verified against official provider pages on . How we verify prices. Written by the Chelure Enterprises Editorial Team.
The idea in one paragraph
An LLM turns text into tokens, converts each token into numbers, and passes them through a neural network that outputs a probability for every possible next token. Generating a reply means choosing a token, appending it and repeating. Everything else, including answering questions and writing code, comes from how large the network is, how much text it was trained on and how it was then tuned.
Four papers that explain how we got here
| Year | Paper | What it contributed |
|---|---|---|
| 2017 | Attention Is All You Need (Vaswani et al.) | Introduced the Transformer, a network based solely on attention mechanisms, dispensing with recurrence and convolutions; it reached 28.4 BLEU on WMT 2014 English-to-German translation |
| 2020 | Scaling Laws for Neural Language Models (Kaplan et al.) | Found that language model loss follows a power law in model size, dataset size and training compute across more than seven orders of magnitude, and that larger models are more sample-efficient |
| 2020 | Language Models are Few-Shot Learners (Brown et al.) | Showed that a 175 billion parameter model could perform tasks from a few examples in the prompt, with no gradient updates or fine-tuning |
| 2022 | Training language models to follow instructions with human feedback (Ouyang et al.) | Fine-tuned GPT-3 with human demonstrations and reinforcement learning from human feedback; a 1.3B InstructGPT model was preferred to the 175B GPT-3 in human evaluations |
How a model is built
- Pre-training: the network learns to predict the next token over a very large text collection. The scaling-laws paper shows why labs build bigger models and train them on more data.
- Instruction tuning: the model is fine-tuned on examples of good responses, as in the InstructGPT work.
- Human-feedback tuning: people rank outputs and the model is trained to prefer the better ones, which Ouyang et al. found improved truthfulness and reduced toxic output.
What you actually pay for and work with
- Tokens. Text is split into tokens, and prices and limits are measured in them. See what is a token in AI.
- Context windows. A model can take only so many tokens at once; the current Claude, GPT-6, Gemini 4 and DeepSeek models list 1M tokens or more in our table.
- Prices. Output tokens cost more than input tokens; see why and compare in the price comparison.
- Prompts. How you ask matters; start with the prompt engineering guide.
What LLMs do badly
They can be wrong with confidence, can be steered by text hidden in the material they read (see prompt injection), and their knowledge stops at a training cutoff: for example Anthropic lists a June 2026 reliable knowledge cutoff for its current models. Treat output as a draft to verify, especially in law, health and finance.
Frequently asked questions
What is a large language model?
A large language model (LLM) is a neural network trained on large amounts of text to predict and generate language. Modern LLMs are built on the Transformer architecture introduced in 2017, read and write text as tokens, and are adapted to follow instructions after their initial training.
How does an LLM generate text?
It reads your prompt as tokens and produces the reply one token at a time, each chosen using the tokens before it. That is why output tokens are generated sequentially and why prompt length and reply length both appear on your bill.
How big are today's models?
The providers' model pages we read do not publish parameter counts for their current models. The best-known published figure is for GPT-3, which Brown et al. (2020) describe as having 175 billion parameters.
Why do models need human feedback?
Ouyang et al. (2022) found that larger language models do not inherently follow user intent better and can produce untruthful, toxic or unhelpful outputs; fine-tuning with human feedback made a 1.3 billion parameter model preferred to the 175 billion parameter GPT-3 in human evaluations.
Do LLMs make mistakes?
Yes. Even the instruction-tuned models in the InstructGPT paper still made simple mistakes, and current models can state things confidently that are wrong. Check important facts against primary sources.
Sources
- Vaswani et al.: Attention Is All You Need (arXiv:1706.03762)
- Kaplan et al.: Scaling Laws for Neural Language Models (arXiv:2001.08361)
- Brown et al.: Language Models are Few-Shot Learners (arXiv:2005.14165)
- Ouyang et al.: Training language models to follow instructions with human feedback (arXiv:2203.02155)
- Anthropic: models overview (knowledge cutoff)
Primary sources opened on November 13, 2026. Prices come from our verified price table (see the methodology).