BETA
A system prompt is an instruction block that sets a model's role, rules, and context before any user messages, steering all of its responses.
The system prompt occupies a privileged position in the conversation: models are post-trained to weight its instructions above conflicting user requests, making it the primary steering mechanism for API applications. Production system prompts define persona, output format, tool usage rules, safety boundaries, and domain context, and routinely run to thousands of tokens.
Every one of those tokens is billed as input on every single request, because the system prompt is resent with each API call. A 3,000-token system prompt across a million requests a month is three billion input tokens of recurring spend — exactly the stable, repeated prefix that prompt caching exists to discount.
System prompts are also an operational asset: they encode much of a product's behavior, deserve version control and eval coverage like code, and need re-validation whenever the underlying model changes.
Last revised 2026-07-05 · All glossary terms → · Live AI model pricing →