Benchmarks › Berkeley Function Calling Leaderboard

Berkeley Function Calling Leaderboard

Older benchmarkBy UC Berkeley Gorilla team ↗Results by Berkeley Function Calling Leaderboard ↗
Choose and call tools across single-turn, multi-turn, search and memory tasks
77.5top score · Claude Opus 4.5
39models ranked
25open-weights models
Weighted overall accuracyhigher is better
Oct 20, 2026last updated

How the test works

1

What it is

Choose and call tools across single-turn, multi-turn, search and memory tasks

  • Kinds of tasksFunction calling, multi-turn interactions, agentic tool use
2

Components

5 parts

  • WeightsNon-live single-turn 10%, Live single-turn 10%, Irrelevance detection 10%, Multi-turn 30%, Agentic (web search and memory) 40%
  • How runs combinePinned v4 code weights non-live, live, irrelevance, multi-turn and agentic categories 10/10/10/30/40%
  • Agent / harnessOfficial BFCL v4 reference implementation at f7cf7359b7ac615a0b294831c5ba2bc95ee4a000 (bfcl-eval 2025.12.17)
3

Score

Weighted overall accuracy

  • Graded byDeterministic AST/function-call checks for single-turn
  • Better isHigher
Sources: gorilla.cs.berkeley.edu, github.com

BFCL v4 leaderboardtop 10 of 39

Full leaderboard with filters →
ModelBFCL v4Per taskPriceSpeedSupportsContext
Claude Opus 4.5Anthropic
77.5unstated
—$5 / $25—200K
Claude Sonnet 4.5Anthropic
73.2unstated
—$3 / $15—1M
Claude Haiku 4.5Anthropic
68.7unstated
—$1 / $5—200K
o3 (2025-04-16)OpenAI
63.1unstated
—$2 / $8—200K
Kimi K2 0711Moonshot AIOpen
59.1unstated
—$0.50 / $2—128K
Gemini 2.5 FlashGoogle
56.2unstated
—$0.30 / $2.50—1M
GPT-5.2 (2025-12-11)OpenAI
55.9unstated
—$1.75 / $14—400K
GPT-5 Mini (2025-08-07)OpenAI
55.5unstated
—$0.25 / $2—400K
GPT-4.1 (2025-04-14)OpenAI
54
—$2 / $8—1M
Qwen3 235B A22B Instruct 2507Alibaba (Qwen)Open
52.2unstated
—$0.09 / $0.55—262K

Versions and boards on Model Menu

BoardModels
BFCL v4Shown above
39
BFCL v4 (past run, prompt mode)
16
BFCL v4 (past run, native function calls)
2

Berkeley Function Calling Leaderboard questions

What is Berkeley Function Calling Leaderboard?›

Choose and call tools across single-turn, multi-turn, search and memory tasks. Created by UC Berkeley Gorilla team.

Which model is best on Berkeley Function Calling Leaderboard?›

Claude Opus 4.5 leads BFCL v4 with 77.5, ahead of Claude Sonnet 4.5 (73.2) and Claude Haiku 4.5 (68.7).

What is the best open-weights model on Berkeley Function Calling Leaderboard?›

Kimi K2 0711 with 59.1, then Qwen3 235B A22B Instruct 2507 (52.2).

How is Berkeley Function Calling Leaderboard scored?›

Weighted overall accuracy; deterministic ast/function-call checks for single-turn; state-transition verification for multi-turn; exact answer-field matching for web search..

Which versions of Berkeley Function Calling Leaderboard are there?›

BFCL v4 (39 models), BFCL v4 (past run, prompt mode) (16 models), BFCL v4 (past run, native function calls) (2 models).