Glossary
The terms used throughout these docs, defined precisely.
| Term | Meaning |
|---|---|
| Token | The unit models read and write, roughly ¾ of an English word. Billing is per million tokens. |
| Input tokens | Everything you send: system prompt, full message history, tool schemas, images. Grows every turn of a conversation. |
| Output tokens | What the model generates. Costs 4–5× input on every frontier model. |
| Context window | The maximum tokens a model can consider at once, input plus output. Exceeding it is an error, not silent truncation. |
| Base URL | The origin a client sends requests to. The single most common misconfiguration — see Base URLs. |
| Messages API | Anthropic's request format. POST /v1/messages, replies in a content[] array. |
| Chat Completions | OpenAI's request format. POST /v1/chat/completions, replies in choices[]. |
| Streaming / SSE | Server-sent events delivering the reply token by token instead of in one body. |
| stop_reason | Why generation ended: end_turn, max_tokens, stop_sequence or tool_use. |
| Prompt caching | Re-reading a stable prefix at a reduced input rate. See Prompt caching. |
| Tool use | The model requesting that you run a function, then continuing with the result. |
| MCP | Model Context Protocol — a standard way to expose tools to an agent. |
| Credits | Prepaid balance spent at per-token rates. Do not expire. |
| Unlimited plan | Flat-rate access for a fixed window, under fair-use limits, instead of per-token billing. |
| Gateway | A service that fronts one or more model providers behind a single API and key. |