Bfcl

BFCL, the Berkeley Function Calling Leaderboard, tests whether a model can use tools correctly. The current version, V4, came out in July 2025. The model is given a list of available functions and a request, and has to pick the right function, fill in the right arguments, and, when nothing fits, call nothing at all.

Correctness is checked mechanically: the call the model produced is compared to the expected one, function by function and argument by argument, and multi-step tasks are checked by the state they leave behind. The overall score is a weighted mix: 40% agentic tasks (searching the web to answer a multi-step question, and storing then recalling facts across a conversation), 30% multi-turn conversations, 10% real requests users submitted to the leaderboard, 10% curated single calls, and 10% "hallucination" cases where the correct answer is to call no tool.

A high score means the model is precise with tool calls in a controlled setting. It says less about your agent, because your tools have their own quirks and your users don't phrase requests like a benchmark. The hallucination slice is the one to watch. A model that calls a tool when it shouldn't is the model that takes an action nobody asked for.

V4 is where BFCL stopped being a single-call test and became an agent test. The earlier versions measured one call at a time, and most top models had saturated them. The agentic and multi-turn parts now decide the ranking.