BFCL

How BFCL works

The benchmark covers simple, multiple, parallel and nested function calls, plus multi-turn interactions. Scoring is primarily done with a deterministic abstract-syntax-tree matcher that compares the model’s invocation against a reference call, alongside executable checks where a real call can be run. V4, released in 2026, extends the suite toward holistic agentic evaluation including web search. The overall score is the unweighted mean of the subcategories.

History and current status

The UC Berkeley Gorilla team introduced BFCL in 2024, and it became the default reference for tool-calling reliability as function calling turned into the foundation of every agent product. Successive versions added abstract-syntax-tree metrics, enterprise and open-source function sets, and multi-turn interaction. V4 in 2026 moved from testing calls in isolation to evaluating agentic behaviour end to end.

What the score does not tell you

Matching against a reference invocation rewards format discipline as much as judgment: a model that makes a sensible but differently structured call can be marked wrong. The unweighted mean across subcategories also means a model can post a respectable overall score while being weak in the one category you depend on, so the subcategory breakdown matters far more than the headline number.

This is the general case, not a quirk of one benchmark. For the four failure modes that decide whether any benchmark number means anything, read are AI benchmarks reliable, and for this row in particular, how benchmark contamination inflates scores.

What BFCL scores actually mean

BFCL scores compress a lot of different abilities into one mean, and the subcategories diverge more than the headline suggests. Simple single-function calls are close to solved for any competent model; parallel and nested calls, and especially multi-turn interactions, are where models separate. Because most of the scoring matches a generated call against a reference invocation using abstract-syntax-tree comparison, BFCL rewards format discipline as much as judgment: a model that picks the right tool but formats the arguments unconventionally is penalised, and a model that produces a perfectly shaped call to the wrong tool can still score on structure. Read the subcategory breakdown, not the unweighted average.

Who reports BFCL, and how to read it

Berkeley runs the leaderboard, which makes it academically independent, and it is updated continuously. Frontier labs rarely lead with BFCL, so if you need to know whether a model will respect your tool schemas, you generally have to go and read the board rather than the launch post. The multi-turn and parallel subcategories are the ones that predict retry costs in a live agent loop.

When to weight BFCL in a model choice

Check BFCL before wiring a model into your own tool definitions, because tool-calling reliability is the single most common cause of expensive agent retries and it is not predicted by coding or reasoning scores. Go straight to the multi-turn and parallel subcategories, which reflect real agent loops, and ignore the simple-call column, which no longer separates anything. Pair it with tau-bench: BFCL tells you whether the calls will be well-formed, tau-bench tells you whether the conversation reaches a correct outcome. For agents that operate a computer rather than an API, use OSWorld 2.0 instead.