Sign in
Benchmarks
Agents
Dated

There is no trustworthy current leaderboard, so the most recent reliable figure is already months old.

BFCL (Berkeley Function Calling)

BFCL, the Berkeley Function Calling Leaderboard, tests whether a model can correctly call software tools: pick the right function and fill in its arguments when asked to actually do something, not just chat about it.

What it measures

Version 4 extends this to full agent tasks like web search and memory. It is the standard scoreboard for the tool-calling that underpins AI agents.

Example task

Given a request like 'book a table for four at 7pm' and a set of available functions, the model must emit the correct function call with the right arguments (date, time, party size). BFCL checks the call against the expected one, so a plausible-looking but wrong argument fails.

Why you should care

If you are building anything that takes actions (querying an API, running code, browsing the web), tool-calling accuracy is the signal that matters most, and BFCL is the most-cited independent measure of it. It is run by a university lab, not the model makers, so the comparison is apples to apples.

How scoring works

Overall accuracy across function-calling categories, evaluated independently by UC Berkeley's Gorilla team against a pinned, reproducible code commit (all model responses are published). Higher is better. Maintained at gorilla.cs.berkeley.edu; the board was last updated in April 2026.

How to read the numbers

Scores average across many call types; the current leaders sit in the low-to-mid 70s, strong open models trail a few points back, and weak models drop into the 50s and below. There is no single human baseline because the task is machine-checkable correctness, not human judgment.

What to watch

The main caveat is freshness: the official board was last updated in April 2026, so the newest frontier models (Claude Opus 4.7 and 4.8, Gemini 3.1, GPT-5.5) are not yet ranked on it, and the current top reflects the prior generation. The headline page is also JavaScript-rendered; the real data lives in a companion CSV at the same site. Tool-calling accuracy on a fixed test can also overstate reliability in messy real-world use.

To go deeper

The benchmark creates the need to know. The catalog explains the ideas behind it:

Leaders as of APR 2026
Claude-Opus-4-5-20251101 (FC)Anthropic
77.47% Overall AccBerkeley-evaluated, reproducible
Claude-Sonnet-4-5-20250929 (FC)Anthropic
73.24% Overall Accabout 4 points behind the leader
Gemini-3-Pro-Preview (Prompt)Google
72.51% Overall Accnear-tie with second place
StatusDated
Board last updated APR 2026, may be behind.
View the leaderboard