What does a BFCL score mean?

A BFCL score is a weighted average across five separately-graded categories, not one uniform test, and a model’s single overall number can hide how lopsided it actually is underneath. On the V4 leaderboard, Grok-4-0709 sits tenth overall at 61.38%, but its Multi-Turn score, the category weighted second-heaviest after Agentic, is only 33.88%, a 27-point gap between the number you’d quote and the category actually dragging it down.

The gap doesn’t disappear at the top either. Claude Opus 4.5, first overall at 77.47%, still posts its weakest score among the five weighted categories on Multi-Turn, at 68.38%, nine points under its overall line. Irrelevance detection, correctly calling nothing when no function fits, is usually the opposite story: most models in the top fifteen score higher there than their overall number, so a low overall score is rarely an over-eager model calling tools it shouldn’t. Which categories BFCL actually weights is the place to check before reading one percentage as a verdict on tool-calling generally.