BFCL-v4

BFCL-v4

Quick answer: BFCL-v4 is the fourth major version of UC Berkeley's Function Calling Leaderboard, introducing holistic agentic evaluation — including a web-search category — on top of the prior versions' AST-verified single- and multi-turn function-calling tests. As of the live leaderboard, Claude Opus 4.5 leads at 77.47% overall accuracy.

At a Glance

What it tests: Function/tool-calling accuracy across simple, multiple, parallel, and multi-turn scenarios (inherited from BFCL-v1–v3), extended with holistic agentic categories including web search.

Why it matters: BFCL is the most widely cited, continuously-updated public leaderboard for tool-calling accuracy, using deterministic AST-based verification rather than LLM-as-judge scoring.

Known limitations: Overall accuracy is an unweighted average across sub-categories with very different difficulty and task counts, so two models with the same headline score can have very different strengths across categories (e.g. multi-turn vs. web search vs. format sensitivity).

What BFCL-v4 Measures

BFCL-v4 builds on the AST-based, deterministic scoring methodology introduced in earlier BFCL versions — parsing a model's generated function call into an abstract syntax tree and comparing it structurally against ground truth, rather than relying on a judge model. Version 4 ("From Tool Use to Agentic Evaluation") extends this into more holistic agentic territory, adding a web-search evaluation category alongside the existing single-turn, parallel, and multi-turn function-calling tests carried over from BFCL-v2 and BFCL-v3.

Models are evaluated in two tracks: FC (native function-calling APIs) and Prompt (text-generation fallback formatted as tool calls), with results reported separately since the two approaches can produce meaningfully different scores for the same underlying model.

Benchmark Specifications

How BFCL-v4 Is Scored

Each generated function call is parsed as an Abstract Syntax Tree and compared against the ground-truth call structurally, producing a deterministic match/no-match result per test case. The headline "Overall Accuracy" is the unweighted average across all sub-categories (including the new web-search category), computed separately for the FC and Prompt evaluation tracks. The public leaderboard is updated periodically as new models are submitted.

State-of-the-Art Results

Scores sourced from the live Berkeley Function Calling Leaderboard, last updated 2026-04-12 on that leaderboard. Results depend on FC vs Prompt evaluation mode — see source for full breakdown by sub-category.

BFCL-v4 on Benchgen

No Benchgen results yet — be the first to run BFCL-v4.

BFCL-v4 vs Other Benchmarks

BFCL remains the standard for deterministic, reproducible function-calling correctness; ToolSandbox better captures stateful, conversational agent behavior that AST matching alone doesn't test.

FAQ

What is BFCL-v4? BFCL-v4 is the fourth version of the Berkeley Function Calling Leaderboard, extending prior AST-verified function-calling evaluation with holistic agentic categories, including web search.

How is BFCL-v4 scored? Overall accuracy is the unweighted average of AST-verified accuracy across sub-categories, reported separately for native function-calling (FC) and prompt-based tool-call generation.

What is a good BFCL-v4 score? As of the live leaderboard, scores above 70% overall accuracy represent frontier-tier function-calling and agentic tool-use performance; the current top model (Claude Opus 4.5) scores 77.47%.

Who created BFCL-v4? Patil et al. at UC Berkeley (Gorilla project), as part of the continuously updated Berkeley Function Calling Leaderboard.

How does BFCL-v4 differ from BFCL-v3? BFCL-v3 focused on single- and multi-turn function-calling accuracy; BFCL-v4 adds a holistic agentic evaluation layer, including web-search tasks, on top of the same AST-verification methodology.

Benchmark definition paraphrased from Patil et al. 2025 and the live Berkeley Function Calling Leaderboard. Last updated 2026-08-04.