A language model is an AI model trained on huge amounts of text to predict how a text continues. That is why it can draft, answer questions, summarise, classify or extract data. Large language models (LLMs), such as Claude, GPT or Gemini, are the engine behind chatbots and AI agents.
In practice
Anthropic's glossary explains that these models are first trained to predict the next word and then refined, with techniques such as reinforcement learning from human feedback, so that they follow instructions. Four ideas help you understand what you are paying for:
| Concept | What it is | Why it matters to you |
|---|---|---|
| Token | The smallest unit of text the model processes: a word, part of a word or a character | You pay per input and output token |
| Context window | What the model can have in front of it when it answers, its "working memory" | Your data is passed in with each request, for example with RAG |
| Temperature | The setting that controls how much answers vary | Even at 0 it is not fully deterministic: the same input can give different outputs |
| Retirement | The date a provider stops serving a model | Requests to a retired model fail; Anthropic gives at least 60 days' notice |
As of October 2026, Anthropic charges $0.10 per million input tokens and $0.50 per million output tokens for Claude Haiku 5.5, and $2 and $10 for Claude Sonnet 5.5 (Anthropic pricing). Anthropic also warns that its models from Claude 4.7 onwards produce about 30% more tokens for the same text, so measure with the model you plan to use. There is a full worked example in how much an AI chatbot costs.
In a small business, the model rarely works alone: it reads an email inside a workflow, pulls the fields out of an invoice or drafts a reply that a person checks. When it also decides which steps to take and uses tools, we are talking about an AI agent.
The EU AI Act gives large generative models as a typical example of a "general-purpose AI model" (recital 99 and Article 3(63)). The obligations for those models fall on their providers (Article 53); if your business uses an AI system in its work, you are a "deployer" (Article 3(4)).
Common mistakes
- Treating it as a database. Anthropic admits that even the most advanced models can generate incorrect text. Important facts come from your documents, not from its memory.
- Using a personal account with customer data. The provider processes that data on your behalf: it is a data processor. Check where the data is processed too: as of October 2026, Anthropic's direct API only offers global or US-only inference.
- Not retesting when you switch models. A new model answers differently. Rerun your set of test cases before you switch. We explain how in automating business processes with AI.