Benchmarks / Berkeley Function Calling Leaderboard (BFCL) V4

Reported by Berkeley Function Calling Leaderboard (BFCL) V4

Berkeley Function Calling Leaderboard (BFCL) V4

Completing tasks over several turns with stateful tools, including when functions or parameters are missing.

Data older than 90 days: last updated 16 Dec 2025

Results dated
16 Dec 2025
Results
109 configurations of 81 models
Unit
% correct
Licence
Apache License 2.0

Berkeley Function Calling Leaderboard (BFCL) V4 in brief

What it measures. Berkeley Function Calling Leaderboard (BFCL) V4 is a third-party benchmark, shown here with credit. Measured here: Completing tasks over several turns with stateful tools, including when functions or parameters are missing.

Latest results. , 81 models ranked.

Top 3 on Multi-turn tasks (% correct, higher is better, each model at its best setting):

  1. xLAM-2-70b-fc-r (Salesforce): 77.4%
  2. xLAM-2-8b-fc-r (Salesforce): 70.0%
  3. xLAM-2-32b-fc-r (Salesforce): 69.5%

How to cite. Berkeley Function Calling Leaderboard (BFCL) V4, via Spring Prompt, results dated 16 December 2025. https://springprompt.com/benchmarks/bfcl These results belong to Berkeley Function Calling Leaderboard (BFCL) V4, not to us. Licence: Apache License 2.0. Credit Berkeley Function Calling Leaderboard (BFCL) V4 first.

Multi-turn tasks: Grok 4.1 Fast (non-reasoning)

Top 15 of 81 results · % correct, higher is better. Choose a model to highlight it.Clear highlight

  1. 1 xLAM-2-70b-fc-rFunction callingSalesforce 77.4%
  2. 2 xLAM-2-8b-fc-rFunction callingSalesforce 70.0%
  3. 3 xLAM-2-32b-fc-rFunction callingSalesforce 69.5%
  4. 4 Claude Opus 4.5Function callingAnthropic 68.4%
  5. 5 GLM-4.6 (reasoning on)Function callingZ.ai 68.0%
  6. 6 Gemini 3 Pro PreviewFunction callingGoogle 63.1%
  7. 7 BitAgent Bounty 8BMode not statedBittensor 62.4%
  8. 8 o3PromptingOpenAI 62.3%
  9. 9 Claude Sonnet 4.5Function callingAnthropic 61.4%
  10. 10 Grok 4.1 Fast (reasoning)Function callingxAI 58.9%
  11. 11 xLAM-2-3b-fc-rFunction callingSalesforce 58.4%
  12. 12 Arch-Agent-32BMode not statedKatanemo 54.3%
  13. 13 Claude Haiku 4.5Function callingAnthropic 53.6%
  14. 14 Nanbeige4-3B-Thinking (2511)Function callingNanbeige 51.1%
  15. 15 Kimi K2 (0711)Function callingMoonshot AI 50.6%
  16. 19 Grok 4.1 Fast (non-reasoning)Function callingxAI 46.8%

Full results

BFCL: multi-turn tasks, % correct, higher is better
#ModelMulti-turn tasks
% correct, higher is better
Price
$ per million tokens, in / out
1 xLAM-2-70b-fc-rFunction callingSalesforce
77.4%
–
2 xLAM-2-8b-fc-rFunction callingSalesforce
70.0%
–
3 xLAM-2-32b-fc-rFunction callingSalesforce
69.5%
–
4 Claude Opus 4.5Function callingAnthropic · best of 2 settings
68.4%
$5 / $25
5 GLM-4.6 (reasoning on)Function callingZ.ai
68.0%
$0.50 / $2
6 Gemini 3 Pro PreviewFunction callingGoogle · best of 2 settings
63.1%
–
7 BitAgent Bounty 8BMode not statedBittensor
62.4%
–
8 o3PromptingOpenAI · best of 2 settings
62.3%
$2 / $8
9 Claude Sonnet 4.5Function callingAnthropic · best of 2 settings
61.4%
$3 / $15
10 Grok 4.1 Fast (reasoning)Function callingxAI
58.9%
–
11 xLAM-2-3b-fc-rFunction callingSalesforce
58.4%
–
12 Arch-Agent-32BMode not statedKatanemo
54.3%
–
13 Claude Haiku 4.5Function callingAnthropic · best of 2 settings
53.6%
$1 / $5
14 Nanbeige4-3B-Thinking (2511)Function callingNanbeige
51.1%
–
15 Kimi K2 (0711)Function callingMoonshot AI
50.6%
$0.57 / $2.30
16 Command A ReasoningFunction callingCohere
50.1%
–
17 Qwen3-32BFunction callingAlibaba · best of 2 settings
47.9%
$0.14 / $0.40
18 Grok 4 (0709)PromptingxAI · best of 2 settings
47.0%
–
19 Grok 4.1 Fast (non-reasoning)Function callingxAI
46.8%
–
20 Qwen3-235B-A22B-Instruct-2507Function callingAlibaba · best of 2 settings
45.4%
$0.15 / $0.75
21 DeepSeek-V3.2-Exp (reasoning on)PromptingDeepSeek · best of 2 settings
44.9%
$0.27 / $0.41
22 GPT-5.2PromptingOpenAI · best of 2 settings
43.8%
$1.75 / $14
23 Qwen3-8BFunction callingAlibaba · best of 2 settings
41.8%
–
23 o4-miniFunction callingOpenAI · best of 2 settings
41.8%
–
25 Nanbeige3.5-Pro-ThinkingFunction callingNanbeige
40.0%
–
26 GPT-4.1Function callingOpenAI · best of 2 settings
38.9%
$2 / $8
27 ToolACE-2-8BFunction callingHuawei Noah's Ark Lab and USTC
38.4%
–
28 Gemini 2.5 FlashFunction callingGoogle · best of 2 settings
36.3%
$0.30 / $2.50
29 xLAM-2-1b-fc-rFunction callingSalesforce
36.0%
–
30 Arch-Agent-3BMode not statedKatanemo
34.9%
–
31 Qwen3-14BFunction callingAlibaba · best of 2 settings
34.8%
$0.10 / $0.22
32 GPT-5 nanoFunction callingOpenAI · best of 2 settings
34.5%
$0.05 / $0.40
33 GPT-4.1 miniFunction callingOpenAI · best of 2 settings
34.1%
$0.40 / $1.60
34 Qwen3-30B-A3B-Instruct-2507Function callingAlibaba · best of 2 settings
30.0%
$0.09 / $0.30
35 Command AFunction callingCohere
29.5%
$2.50 / $10
36 GPT-5 miniFunction callingOpenAI · best of 2 settings
27.5%
$0.25 / $2
37 Arch-Agent-1.5BMode not statedKatanemo
26.6%
–
38 Hammer2.1-7BFunction callingMadeAgents
23.9%
–
39 GPT-4.1 nanoFunction callingOpenAI · best of 2 settings
23.6%
$0.10 / $0.40
40 Qwen3-4B-Instruct-2507Function callingAlibaba · best of 2 settings
22.1%
–
41 Llama 3.3 70B InstructFunction callingMeta
21.5%
$0.71 / $0.71
42 Llama 4 Maverick 17B 128E Instruct (FP8)Function callingMeta
20.3%
–
43 Hammer2.1-3BFunction callingMadeAgents
16.5%
–
44 Hammer2.1-1.5BFunction callingMadeAgents
15.6%
–
45 Mistral Small 3.2 24BPromptingMistral AI · best of 2 settings
14.8%
$0.094 / $0.25
46 Mistral Large 2.1 (2411)Function callingMistral AI · best of 2 settings
14.1%
–
47 Gemini 2.5 Flash-LiteFunction callingGoogle · best of 2 settings
13.5%
$0.10 / $0.40
48 Llama 3.1 8B InstructPromptingMeta
11.1%
$0.05 / $0.08
49 Qwen3-1.7BFunction callingAlibaba
11.0%
–
50 Gemma 3 27BPromptingGoogle
10.8%
$0.08 / $0.35
50 Mistral Medium 3Function callingMistral AI · best of 2 settings
10.8%
$0.40 / $2
52 CoALM 70BMode not statedUIUC and Oumi
10.6%
–
53 Llama 4 ScoutFunction callingMeta
9.0%
$0.10 / $0.30
54 Command R7B (12-2024)Function callingCohere
8.3%
$0.037 / $0.15
55 CoALM 8BMode not statedUIUC and Oumi
8.0%
–
56 Mistral NemoFunction callingMistral AI · best of 2 settings
7.8%
$0.029 / $0.03
57 Granite 3.1 8B InstructFunction callingIBM
7.5%
–
58 Granite 3.2 8B InstructFunction callingIBM
7.4%
–
59 Falcon3 10B InstructFunction callingTII
6.5%
–
60 Gemma 3 12BPromptingGoogle
5.8%
$0.05 / $0.15
61 Granite 20B FunctionCallingFunction callingIBM
5.4%
–
62 Falcon3 7B InstructFunction callingTII
5.0%
–
63 Llama 3.2 3B InstructFunction callingMeta
4.0%
$0.05 / $0.33
64 Phi-4PromptingMicrosoft
3.9%
$0.07 / $0.14
64 MiniCPM3-4BFunction callingOpenBMB · best of 2 settings
3.9%
–
66 Qwen3-0.6BFunction callingAlibaba · best of 2 settings
3.6%
–
67 Hammer2.1-0.5BFunction callingMadeAgents
2.9%
–
67 RZN-TPromptingPhronetic AI
2.9%
–
69 Bielik 11B v2.3 InstructPromptingSpeakLeash and ACK Cyfronet AGH
2.6%
–
70 Granite 4.0 350MFunction callingIBM
2.5%
–
71 Nova 2 LiteFunction callingAmazon
2.1%
$0.30 / $2.50
72 Nova Pro 1.0Function callingAmazon
1.9%
$0.80 / $3.20
73 Nova Micro 1.0Function callingAmazon
1.4%
$0.035 / $0.14
74 Falcon3 3B InstructFunction callingTII
1.0%
–
75 Gemma 3 4BPromptingGoogle
0.4%
$0.04 / $0.05
75 Palmyra X 004Function callingWriter
0.4%
–
77 Gemma 3 1BPromptingGoogle
0.0%
–
77 Llama 3.2 1B InstructFunction callingMeta
0.0%
$0.027 / $0.20
77 Ministral 8B (2410)Function callingMistral AI
0.0%
–
77 Llama 3.1 Nemotron Ultra 253B v1Function callingNVIDIA
0.0%
–
77 Falcon3 1B InstructFunction callingTII
0.0%
–

Swipe the table sideways for more columns.

Source: Berkeley Function Calling Leaderboard (BFCL) V4, via Spring Prompt, springprompt.com/benchmarks/bfcl. Test your own prompt on these models →

Ranks follow the score as shown, so equal numbers share a rank. Each model is shown at its best setting; show every setting. Results as published by Berkeley Function Calling Leaderboard (BFCL) V4; we do not re-run them.

Progress over time

Multi-turn tasks, plotted at each model's release date, higher is better. Each model counts once, at its best setting. Scores are today's, not launch day's; 38 models without a known release date are left off. Hover or tap a mark for its model.

Best score so far

Every model at its release date. The line moves each time a new release beats every earlier one, and each new best is named where there is room.

-20%0%20%40%60%80% Jul 20242025Jul2026Jul Mistral NemoLlama 3.1 8B InstructMistral Large 2.1 (2411)Llama 3.3 70B InstructGPT-4.1o3GLM-4.6Claude Opus 4.5

Best score by lab

Each lab's best so far, stepping whenever one of its releases beats its own earlier models. The 6 labs with the best scores today; 7 others not drawn.

AnthropicZ.aiGoogleOpenAIxAIMoonshot AI

0%20%40%60%80% Jul 20242025Jul2026Jul AnthropicZ.aiGoogleOpenAIxAIMoonshot AI

What it measures

Completing tasks over several turns with stateful tools, including when functions or parameters are missing.

What it does not measure

Not open-ended agent work: the tools and tasks are BFCL's simulated APIs.

Berkeley Function Calling Leaderboard (BFCL) V4 by the UC Berkeley Gorilla team, Apache License 2.0.