Function calling
An LLM API feature that lets the model emit a structured JSON call to a developer-defined function — the model picks the function name and arguments; the runtime executes the call.
Function calling is the API primitive that powers tool use. You declare a function schema (name, description, parameter types) when you make the LLM call; the model can choose to emit a structured invocation matching that schema instead of free-text. Your code then executes the call and feeds the result back.
OpenAI introduced the canonical API in 2023; Anthropic, Google, and most open-weights models followed. In 2026 the feature is table-stakes — what differs is reliability (can the model pick the right function out of 50?), schema flexibility, and parallel calling (emitting multiple function calls in one turn).
Function calling is the engine inside every modern agent. The agent's system prompt declares dozens of functions; the model picks which to call; the loop continues until done. Whether the term you read is "function calling", "tool use", or "tool calling", the mechanic is the same.
Frequently asked
Is function calling the same as tool use?+
Function calling is the API mechanic; tool use is the agent capability. The model "function-calls" a tool; the loop is "tool use". The terms are often used interchangeably.
How many functions can an LLM reliably call from?+
Frontier models in 2026 handle 20–30 functions well. Past 50, accuracy drops sharply and most teams either split into specialized agents or move tool selection into a retrieval step before the LLM call.
Can the model call multiple functions in one turn?+
Yes — most 2026 frontier APIs support parallel function calling. The model emits a list of calls; the runtime executes them concurrently and returns results together. Cuts agent latency by 30–60% on multi-tool steps.