The Berkeley Function-Calling Leaderboard (BFCL, also known as the Berkeley Tool Calling Leaderboard) is used to showcase the ability of large language models to accurately call functions (i.e., tools).
Purpose: This leaderboard evaluates the ability of large language models to accurately call functions (tools). BFCL V4 is described as spanning from tool use to large language model agentic evaluation, and the leaderboard is composed of real-world data and is updated regularly.
Main capabilities and evaluation content: The leaderboard evaluates model function-calling ability based on leaderboard data. The evaluation distinguishes between FC (native support for function/tool calling) and Prompt (workarounds for function calling achieved through the model's regular text generation ability). The metrics shown on the leaderboard include cost, latency, and overall accuracy. Cost is calculated as the estimated cost for the entire benchmark, in US dollars; latency is in seconds; overall accuracy is the unweighted average of all subcategories. Format sensitivity test cases only support Prompt (non-FC) models.
Versions and background: BFCL-v1 introduced AST as an evaluation metric; BFCL-v2 introduced functions contributed by enterprises and OSS; BFCL-v3 introduced multi-turn interaction; BFCL-v4 introduced holistic agentic evaluation.
How to use: Users can explore the leaderboard to view model function-calling ability; they can click column headers to sort; in the charts, they can select or deselect the models to compare. The page also provides a Function Calling Demo, where prompts and functions can be entered and outputs viewed; the outputs include both the actual code format and the OpenAI-compatible format; the OpenAI-compatible format output is only available when the actual code output is syntactically valid and parseable, and the page also provides several examples for trial. To add a model or contribute test cases, contact via Discord.
Reproducibility and resources: Models are evaluated using commit f7cf735, and all model responses are available; to reproduce the results, check out the codebase at that checkpoint, or install the PyPI package with pip install bfcl-eval==2025.12.17. The code and data can be viewed.
Limitations: For more information about the dataset and methodology, refer to its blog (BFCL-v1, v2, v3, v4); format sensitivity test cases only support Prompt (non-FC) models; the OpenAI-compatible format output is only available when the actual code output is syntactically valid and parseable; adding a modelor contributing test cases requires contacting via Discord.
