Which agent benchmark should I use for tool calling?

For tool calling specifically, use BFCL: it’s built to grade function selection and argument formation directly, rather than folding tool use into a broader end-to-end task the way conversational or coding benchmarks do. The current version, V4, checks whether a model picks the right function, fills in the right arguments, and calls nothing when nothing fits, weighted 40% agentic tasks (web search and multi-step recall), 30% multi-turn conversations, 10% real user-submitted requests, 10% curated single calls, and 10% cases where the correct call is none at all. That last slice is worth watching on its own: a model that scores well everywhere else but calls a tool it shouldn’t is a model that takes an action nobody asked for.

BFCL still grades against its own fixed function list and its own scoring harness, not your tool definitions or your users’ phrasing. A high score says the model is precise with tool calls in a controlled setting; how those calls hold up once a real tool can time out or rate-limit is a separate, later question BFCL doesn’t answer.