The short answer
A large language model (LLM) is an AI model trained on huge amounts of text to predict the next piece of text, which lets it draft, summarise, translate, answer questions and write code. Google's Machine Learning Glossary defines it as a language model with a very high number of parameters. In November 2026, ChatGPT runs on GPT-6 Sol and GPT-6 Luna, Claude on Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5, and Gemini's Google AI Pro and Ultra plans on Gemini 4 Argon.
What a large language model is, in plain English
A large language model is a program that has read an enormous amount of text and learned the patterns in it so well that it can continue any text you give it. You type a question, and the model produces the words most likely to make a good answer, one small piece at a time. Those pieces are called tokens.
"Large" refers to the number of parameters, the internal settings the model adjusts while it learns. Google's Machine Learning Glossary puts it simply: at a minimum, a language model with a very high number of parameters. Anthropic's glossary adds the useful part: LLMs are trained on vast amounts of text and can generate human-like text, answer questions and summarise information. For scale, DeepSeek publishes the size of its open model DeepSeek-V4-Pro: 1.6 trillion parameters, of which 49 billion are active for each token.
The European Union uses a related legal term. The EU AI Act calls the biggest of these models general-purpose AI models: models trained with a large amount of data using self-supervision at scale that display significant generality and can perform a wide range of distinct tasks. Most LLMs behind today's assistants fall into that category.
How an LLM learns: pretraining, then fine-tuning
Building an LLM happens in two broad stages.
- Pretraining. The model reads a very large body of text and learns to predict the next word from the words before it. Anthropic notes that a model at this stage is not good at answering questions or following instructions yet.
- Fine-tuning and feedback. The maker then trains it further on examples of helpful answers and on human ratings, a method called reinforcement learning from human feedback (RLHF). This is what turns a text predictor into an assistant that follows instructions. Anthropic says Claude was trained this way. You can read more in what is fine-tuning.
The result is a model with a frozen snapshot of knowledge from its training data. Anything newer has to come from somewhere else: a web search, a file you upload, or a company knowledge base connected through retrieval-augmented generation (RAG).
Why LLMs matter at work
An LLM is the engine; the app you open is the car. ChatGPT, Claude, Gemini, Microsoft Copilot and Perplexity all put an LLM behind a chat box, then add tools such as web search, file reading, image creation and connections to your email or calendar. Three practical consequences follow.
- The model you get depends on your plan. In ChatGPT, Free and Go users get GPT-6 Luna, while Plus, Pro, Business and Enterprise users get GPT-6 Sol in the Chat tab (since 7 October 2026). In Gemini, Google AI Pro and Ultra get Gemini 4 Argon, Google AI Plus gets a Flash model, and free accounts use Auto, which since 9 October 2026 sends most prompts to the smaller Flash-Lite model and harder ones to Flash or Pro. Two colleagues typing the same prompt can be talking to different models.
- LLMs are strongest with language work. Drafting, rewriting, summarising, translating, classifying feedback and explaining code are the jobs where they save the most time.
- LLMs predict; they do not look things up. Unless the tool searches the web or reads your files, the answer comes from patterns in training data. That is why a fluent answer can still be wrong, a problem covered in what is hallucination.
The LLMs behind the tools you use, November 2026
| Assistant | Maker | Current models (by plan) |
|---|---|---|
| ChatGPT | OpenAI | GPT-6 Luna (Free, Go), GPT-6 Sol (paid chat), GPT-6.1 Sol and GPT-6.1 Terra (ChatGPT Work and Codex on Plus and higher plans) |
| Claude | Anthropic | Claude Sonnet 5.5 and Haiku 5.5 (all plans), Claude Opus 5.5 (Pro and above) |
| Gemini | Auto, mostly Flash-Lite (free), Flash (Google AI Plus), Gemini 4 Argon (Google AI Pro and Ultra) | |
| Meta AI | Meta | Muse Spark, with Muse Spark 1.1 in Thinking mode |
| DeepSeek | DeepSeek | DeepSeek-V4.1-Flash-1017, and DeepSeek-V4-Pro through Expert Mode |
| Vibe (formerly Le Chat) | Mistral AI | Mistral's own models; Mistral made Mistral Medium 3.5, a 128B open-weight model, the default in Le Chat in May 2026 |
Some of these models are open-weight, meaning anyone can download and run them: DeepSeek publishes DeepSeek-V4-Pro on Hugging Face under the MIT licence. Others, such as GPT-6 Sol and Claude Opus 5.5, are only available through the maker's apps and API. If you are choosing between assistants rather than models, start with ChatGPT vs Claude or ChatGPT vs Gemini, and see how to choose the right AI model.
Try it: see what an LLM is good and bad at in ten minutes
Open any assistant you have, even on a free plan, and run these three tests. They teach more than any definition.
- A language task. Paste a messy paragraph from a real email and ask for a clearer version in under 80 words. This is the job LLMs do best, and the result is usually good on the first try.
- A recent fact with search off. Ask about something that happened last week without asking it to search. Watch whether it admits it does not know or guesses. Then ask again and tell it to search the web, and compare.
- The same prompt in two tools. Paste the first prompt into a second assistant, for example ChatGPT and Claude. The differences in tone and structure show you that each LLM has its own habits, which is why the comparison tool exists.
For a structured way to write those prompts, read how to write a good prompt.
Common misconceptions about LLMs
- "It looks things up in a database." It does not, unless the app adds a search or retrieval tool. The model itself generates text from learned patterns.
- "It remembers everything I told it last month." The model does not learn from your chat in real time. Apps add separate memory features, and what the model can see in one conversation is limited by its context window.
- "Bigger is always better." Smaller models such as Claude Haiku 5.5 or GPT-6 Luna are faster and often good enough for routine drafting. Use the larger one when the task needs careful reasoning.
- "An LLM and a chatbot are the same thing." The chatbot is the product; the LLM is one part of it. The same LLM can sit behind a chat app, a coding tool or an AI agent.
- "My chats never reach the model maker." On many personal plans they can be used for training unless you switch it off. In ChatGPT the setting is Improve the model for everyone in Data controls. See how to keep your data private in AI tools.
Related terms
- Generative AI is the wider family of AI that creates new content, of which LLMs are the text branch.
- Token is the unit of text an LLM reads and writes.
- Prompt is what you type to steer the model.
- Multimodal AI covers models that also handle images, audio or video.
- Open-source model covers models whose weights you can download and run yourself.
- Reasoning model: LLMs that work through a problem step by step before answering.
Prompts to copy
Test what the model knows and does not know
Before you answer, tell me how current your knowledge is and whether you can search the web in this chat. Then answer: [question about a recent event in your industry]. If you are not sure, say so instead of guessing.
You get: An answer that shows whether the assistant is working from training data or live search, and how it handles uncertainty.
Rewrite a real email
Rewrite this email so it is clear, polite and under 80 words. Keep every date, name and number exactly as written. Email: [paste email]
You get: A shorter, clearer version of your email, which shows the kind of language work LLMs do best.