How large language models actually work, what they cost, and where they help or fail.

A large language model (LLM) is an AI system trained on huge amounts of text to predict what words come next. That simple mechanism, scaled up, lets it write, summarize, translate, answer questions, and reason through problems. ChatGPT, Claude, and Gemini are all interfaces to LLMs.

An AI model breaks your input into tokens (word pieces), then predicts the most likely next token, adds it, and repeats, one token at a time, until the answer is complete. Every word it produces is a probability-weighted guess informed by patterns it learned during training.

Tokens are the chunks of text an AI model reads and writes. A token is roughly three-quarters of a word in English, so 1,000 tokens is about 750 words. They matter because AI pricing and limits are measured in tokens, not words or characters.

A context window is the maximum amount of text an AI model can hold in mind at once, counted in tokens. It includes your prompt, any documents you attach, the conversation so far, and the answer being written. Once you exceed it, the model loses track of the earliest content.

AI models hallucinate because they generate the most plausible-sounding next words, not verified facts. When they don't know something, they don't stay silent. They produce a confident guess that fits the pattern of a real answer, which is why a made-up citation or statistic can look entirely legitimate.

A prompt is the instruction you give an AI model. Prompt engineering is the practice of writing that instruction well: being specific about the task, giving context and examples, stating the format you want, and setting constraints. Better prompts produce dramatically better output from the same model.

RAG (retrieval-augmented generation) feeds a model your data at question time, so it answers from your documents without changing the model. Fine-tuning retrains a model on your examples so it internalizes a style or task. RAG is for knowledge, fine-tuning is for behavior.

Two costs stack up: the usage cost of the AI model (billed per token) and the cost of building the product around it. As of July 2026, mid-tier models run roughly $2 to $3 per million input tokens and $10 to $15 per million output tokens, with budget models under $0.50 and frontier models higher.

It depends on the job. As of July 2026 the main families are OpenAI's GPT-5 models, Anthropic's Claude models, and Google's Gemini models, each with a small, fast, cheap tier and a larger, more capable one. Match the tier to the task instead of always reaching for the most powerful.

An AI agent is software that uses an AI model to pursue a goal across multiple steps on its own: it plans, takes actions using tools (searching, calling an API, updating a record), checks the result, and adjusts. A plain chatbot waits for each instruction. An agent keeps going until the job is done.

An AI tool does one task when you ask: draft this, summarize that, answer a question. An AI agent chains many steps toward a goal and acts on its own, deciding what to do next and using other tools to get there. Tools assist a person. Agents complete a workflow.

It depends on the provider and plan. As of July 2026, the major AI providers say they don't train on data sent through their business and API tiers by default, though free consumer tiers sometimes may. The safe move is to read the specific terms for the plan you're on before sending anything sensitive.

Not wholesale, and that's usually the wrong goal. As of 2026, AI reliably takes over repetitive, well-defined tasks and speeds up skilled work, but it still needs people to set direction, judge quality, handle exceptions, and own the outcome. The teams that win use AI to do more, not to shrink.

Multimodal AI handles more than text. A multimodal model can read images, listen to audio, watch video, and respond across those formats, not just words. As of 2026 the leading models are multimodal by default, so you can show one a photo, a chart, or a screenshot and ask about it.

An embedding turns a piece of text (or an image) into a list of numbers that captures its meaning, so a computer can measure how similar two things are. Content about close ideas ends up with similar numbers. It's the math that powers semantic search and recommendations.