AI Glossary: 50 Terms Explained in Plain Words (2026)

Updated: November 2026 · Prices checked: November 2026

AI news is full of words that sound technical and mean something simple. You do not need to know how a model works inside to use it well. You do need about 50 words to read a pricing page, compare tools and not get sold something you do not need. Here they are, short and grouped by topic.

The basics (1–10)

1. Artificial intelligence (AI). Software that does tasks we used to think needed a person: writing, reading, recognizing images, deciding. Example: a tool that sorts your inbox by topic.
2. Generative AI. AI that creates new content: text, images, audio, video, code. ChatGPT writing an email is generative AI. A spam filter is not.
3. Large language model (LLM). The engine behind chatbots. A model trained on huge amounts of text to predict the next piece of text. Claude, GPT and Gemini are families of LLMs.
4. Model. One specific trained version, with a name and a price. Vendors release new models every few months. The app stays the same; the model inside changes.
5. Chatbot. The app you type into. ChatGPT, Claude and Gemini are chatbots running on their own models.
6. Prompt. What you type to the model: the question, instructions and material. Better prompts give better answers.
7. Token. The unit AI reads and bills in. A piece of a word. OpenAI’s rule of thumb for English: 1 token is about 4 characters or 0.75 words. Other languages, including Bulgarian, usually need more tokens for the same text.
8. Context window. How much text the model can see at once: your prompt, files and the chat so far. Claude’s plans list up to 1M tokens, depending on the model. Past the limit, older parts drop out or the request fails.
9. Parameters. The internal numbers a model learned in training. More parameters usually means more capable and more expensive to run. Sizes are given in billions (B) or trillions (T).
10. Training data. The text, images and code a model learned from. It decides what the model knows and where it is weak.

The maths: a 2,000-word document is roughly 2,700 tokens in English. Anthropic lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens. Summarize 100 such documents a month, with 300-word summaries of about 400 tokens each. Input: 270,000 tokens, about $0.54. Output: 40,000 tokens, about $0.40. Total: about $0.94, or roughly €0.80. The exact count depends on the model’s tokenizer, so treat it as a ballpark.

How models behave (11–18)

11. Inference. Running a trained model to get an answer. Every message you send is an inference, and that is what you pay for.
12. Hallucination. When the model states something false with confidence: a made-up court case, a wrong price, a fake source. Always check facts that matter.
13. Knowledge cutoff. The date after which the model has no training data. Without web search, it does not know newer events or prices.
14. Temperature. A setting for randomness. Low gives steady, predictable answers. High gives more variety. Mostly visible in API tools, not in chat apps.
15. Reasoning model. A model that works through a problem step by step before it answers. Slower and pricier, better at maths, planning and code.
16. Multimodal. A model that handles more than text: images, audio, PDFs, video. Example: you photograph a receipt and ask for the total.
17. System prompt. Standing instructions the model follows in every reply, set by the app or by you. Custom GPTs, Gems and projects use them to keep your tone and rules.
18. Memory. A feature that lets the chatbot remember facts about you across chats. Handy, and worth checking in settings so you know what it keeps. Anthropic has shown how memory changes how well agents perform.

Customizing and building (19–27)

19. API. A way for software to talk to a model directly, without the chat app. You pay per token instead of a monthly plan. Used to build tools, bots and automations.
20. Fine-tuning. Extra training on your own examples to change how a model writes or behaves. Useful for a fixed style or format. Rarely the first thing a small business needs.
21. RAG (retrieval-augmented generation). The model looks up relevant pieces of your documents and answers from them. Example: a support bot that answers from your help pages.
22. Embeddings. Text turned into lists of numbers that capture meaning, so software can find similar texts. The search step behind most RAG systems.
23. Vector database. A database built to store embeddings and find the closest matches fast.
24. Prompt caching. The vendor stores a long, repeated part of your prompt so later calls are cheaper and faster. Anthropic charges cache reads at a fraction of the normal input price.
25. Batch API. You send many requests to process later instead of instantly. Anthropic lists batch requests at 50% off. Good for overnight jobs.
26. Open-weights model. A model whose trained weights you can download and run yourself. Licenses differ, so check what commercial use is allowed.
27. Local model. A model running on your own computer or server. Your data stays with you. See how to run a 27B Qwen model on a laptop with LM Studio.

Three terms in this group get mixed up all the time. Here is how they compare:

Agents and tools (28–36)

28. AI agent. A model that does not only answer but acts: it plans steps, uses tools, checks results and keeps going until a task is done. Example: it reads new orders and updates your spreadsheet.
29. Tool use (function calling). The model asks your software to do something, such as search, calculate or send an email, and uses the result.
30. MCP (Model Context Protocol). An open standard for connecting AI apps to outside data and tools. Its official docs call it a USB-C port for AI. One connector works with many AI apps instead of one custom integration each.
31. Computer use. An agent that controls a screen, mouse and keyboard to work in apps that have no API.
32. Automation (workflow). A fixed chain of steps, such as “new form entry, then AI summary, then email”. More predictable than an agent. Tools: Make, Zapier, n8n.
33. Guardrails. Rules and limits that keep an AI system inside safe bounds: what it may access, spend or send. Read how to keep agents in line.
34. Human in the loop. A person approves key steps before the AI acts. The simplest safety rule for anything that touches money or customers.
35. Sandbox. A closed environment where an agent can run code or test changes without touching your real systems.
36. Prompt injection. Hidden instructions in a web page, email or file that trick an AI into doing something you did not ask. A real risk for agents. One test showed a single AGENTS.md file hijacking Codex.

Money and limits (37–44)

37. Subscription. A fixed monthly price for the chat app, with usage limits. ChatGPT Plus and Claude Pro are $20 a month, about €17.
38. Pay-per-use (API pricing). You pay for the tokens you use. Cheap for small volumes, open-ended for big ones.
39. Price per million tokens. The standard way APIs list prices, often written as $/MTok. Compare models with it, like Meta Muse Ember at $0.10 per million tokens for repeat agent work.
40. Input and output tokens. Input is what you send, output is what the model writes. Output usually costs several times more, so long answers cost more than long questions.
41. Usage limit (rate limit). A cap on how much you can use per hour, day or week. Claude counts usage in 5-hour windows and per week.
42. API credits. Prepaid API balance. Some top plans include them, for example $100 a month with Claude Max.
43. Latency. How long you wait for the answer. Matters for live chat and voice; less for reports.
44. Benchmark. A standard test used to compare models. Useful as a hint, never as proof that a model suits your task.

Model types and sizes (45–50)

45. Frontier model. The most capable models available at a given time. Usually the most expensive too.
46. Small model. A compact, cheaper model for simple jobs such as sorting, tagging and short replies. Often good enough and much faster.
47. Mixture of experts (MoE). A big model where only part of it works on each token. Example: LongCat-2.5 has 1.6T parameters but only 48B work per token. You get big-model quality at lower running cost.
48. Distillation. Training a small model to copy a big one. That is how many cheap, fast models are made.
49. Quantization. Storing a model’s numbers with less precision so it fits on smaller hardware. A little quality loss, a lot less memory.
50. Open source vs closed. Closed models (most of GPT, Claude, Gemini) run only on the vendor’s servers. Open models can be downloaded. “Open source” in AI often means open weights, not open training data.

Read also: AI Agents Explained: What They Can Do, Costs and Risks (2026) · MCP Explained for Non-Programmers: Connect AI to Your Tools · AI Customer Support That Doesn’t Annoy Customers (2026)

Sources: OpenAI: What are tokens · Anthropic: Glossary · Anthropic: API pricing · Model Context Protocol: Introduction · Claude pricing

II

About the author

Ivan Ivanov

Founder ofHawkValue

If it doesn’t save you time or money, you won’t see it here

• Facts, not hypeEvery price and number is checked against the official source
• Always the real priceYou see what it costs before you click
• I say when it’s weakEven when it’s OpenAI or Google

All articles by Ivan →