BFCL Overall
Berkeley Function Calling Leaderboard overall accuracy.
As of 25 Oct 2026, Claude Opus 4.5 leads BFCL Overall on BenchLeader with 77.5%, ahead of Claude Sonnet 4.5 at 73.2%, across 83 model configurations with a published result.
- Published by
- Berkeley Function Calling Leaderboard
- Category
- Agents & tools
- Index weight
- 1.0
- Models
- 83
- Data as of
- 25 Oct 2026
Apache 2.0 — ShishirPatil/gorilla (BFCL).
What the test looks like
The Berkeley Function Calling Leaderboard tests whether a model calls tools correctly: choosing the right function, filling arguments, handling multiple or parallel calls, and knowing when no call is needed, in single and multi-turn settings.
How it is scored
Overall accuracy across the categories, evaluated by running or checking the calls.
What to keep in mind
Some results are submitted by developers; prompting versus native function-calling modes differ. It measures API-style tool use, not open-ended agency.
- 1Claude Opus 4.577.5%
- 2Claude Sonnet 4.573.2%
- 3Gemini 3 Pro72.5%
- 4GLM-4.6 (thinking)72.4%
- 5Grok 4.1 Fast (thinking)69.6%
- 6Claude Haiku 4.568.7%
- 7o363.0%
- 8Grok 463.0%
- 9Kimi K259.1%
- 10Grok 4.1 Fast (none)58.3%
- 11Command A Reasoning (thinking)57.1%
- 12DeepSeek V3.2 (thinking)56.7%
- 13Gemini 2.5 Flash56.2%
- 14GPT-5.255.9%
- 15GPT-5 mini55.5%
| # | ||||
|---|---|---|---|---|
| 1 |  Claude Opus 4.5 | 77.5% | 57.9 | $86.55 |
| 2 | Claude Sonnet 4.5 | 73.2% | 53.7 | $43.73 |
| 3 |  Gemini 3 Pro | 72.5% | 60.4 | $298.47 |
| 4 |  GLM-4.6thinkingopen ↗ | 72.4% | 49.1 | $4.64 |
| 5 |  Grok 4.1 Fastthinking | 69.6% | 54.7 | $17.26 |
| 6 | Claude Haiku 4.5 | 68.7% | 45.3 | $14.23 |
| 7 |  o3 | 63.0% | 59.5 | $234.64 |
| 8 | Grok 4 | 63.0% | 58.0 | $348.19 |
| 9 |  Kimi K2open ↗ | 59.1% | 50.4 | $6.19 |
| 10 | Grok 4.1 Fastno reasoning | 58.3% | 42.5 | $16.27 |
| 11 |  Command A Reasoningthinkingopen | 57.1% | – | $3.04 |
| 12 |  DeepSeek V3.2thinkingopen ↗ | 56.7% | 55.5 | $57.75 |
| 13 | Gemini 2.5 Flash | 56.2% | 52.7 | $26.36 |
| 14 | GPT-5.2 | 55.9% | 58.0 | $85.65 |
| 15 | GPT-5 mini | 55.5% | 56.5 | $22.18 |
| 16 |  Xlam 2 32B Fc R | 54.7% | – | $6.00 |
| 17 | DeepSeek V3.2open ↗ | 54.1% | 56.6 | $6.71 |
| 18 | GPT-4.1 | 54.0% | 48.3 | $100.75 |
| 19 | o4-mini | 53.2% | 56.9 | $81.91 |
| 20 | Xlam 2 70B Fc R | 53.1% | – | $25.10 |
| 21 |  Qwen3 235B A22B 2507open ↗ | 52.1% | 50.4 | $3.12 |
| 22 | GPT-5 nano | 51.5% | 55.9 | $8.79 |
| 23 | Nanbeige4 3Bthinking | 51.4% | – | $14.14 |
| 24 | GPT-4.1 mini | 50.5% | 45.0 | $19.25 |
| 25 | Qwen3 32Bopen ↗ | 48.7% | 51.5 | $153.08 |
| 26 | Nanbeige3.5 Prothinking | 47.7% | – | $23.46 |
| 27 | Xlam 2 8B Fc R | 46.7% | – | $20.92 |
| 28 | Command Aopen ↗ | 46.5% | 41.8 | $91.37 |
| 29 |  Bitagent Bounty 8B | 46.2% | – | $18.02 |
| 30 | Arch Agent 32B | 45.4% | – | $8.87 |
| 31 | Qwen3 8Bopen ↗ | 42.6% | – | $43.32 |
| 32 |  Toolace 2 8B | 42.4% | – | $24.43 |
| 33 | Qwen3 30B A3B 2507open ↗ | 41.4% | 45.6 | $5.62 |
| 34 | Xlam 2 3B Fc R | 41.2% | – | $3.36 |
| 35 | Qwen3 14Bopen ↗ | 41.0% | 46.9 | $3.38 |
| 36 |  Mistral Large 2.1open | 38.4% | 42.5 | $115.98 |
| 37 | Mistral Medium 3 | 37.7% | 44.1 | $36.51 |
| 38 |  Llama 4 Maverickopen ↗ | 37.3% | 42.8 | $18.25 |
| 39 | Mistral Small 3.2open ↗ | 37.1% | 41.6 | $5.20 |
| 40 | Gemini 2.5 Flash-Lite | 36.9% | 48.3 | $7.55 |
| 41 | Qwen3-4B-Instruct-2507open | 35.7% | 40.9 | $6.37 |
| 42 | Arch Agent 3B | 35.4% | – | $3.70 |
| 43 | GPT-4.1 nano | 33.0% | 38.0 | $5.66 |
| 44 | Arch Agent 1 5B | 32.1% | – | $2.45 |
| 45 | Command R7b | 32.1% | – | $1.50 |
| 46 | Llama 3.3 70Bopen ↗ | 31.9% | 40.7 | $29.54 |
| 47 | Hammer2 1 7B | 31.7% | – | $4.99 |
| 48 | Xlam 2 1B Fc R | 30.4% | – | $2.79 |
| 49 | Gemma 3 12Bopen ↗ | 30.4% | 36.9 | $10.77 |
| 50 | Hammer2 1 3B | 29.7% | – | $10.89 |
| 51 | Gemma 3 27Bopen ↗ | 29.5% | 38.9 | $11.82 |
| 52 |  Phi-4open ↗ | 28.8% | 38.9 | $8.72 |
| 53 | Qwen3-1.7Bopen ↗ | 28.4% | – | $4.33 |
| 54 | Llama 4 Scoutopen ↗ | 28.1% | 39.7 | $24.68 |
| 55 | Coalm 70B | 28.0% | – | $19.89 |
| 56 | Hammer2 1.1 5B | 27.9% | – | $6.83 |
| 57 |  Palmyra X 004 | 27.9% | – | $178.15 |
| 58 | Open Mistral Nemoopen | 27.6% | – | $8.12 |
| 59 |  Nova 2 Lite | 27.1% | 48.9 | $78.19 |
| 60 |  Granite 3.1 8B | 27.1% | 41.8 | $9.32 |
More agents & tools benchmarks · All LLM benchmarks
Cite as: BenchLeader, “BFCL Overall leaderboard”, data as of 25 Oct 2026, https://www.benchleader.com/benchmarks/bfcl_overall.
BFCL Overall: questions
- What does BFCL Overall measure?
- The Berkeley Function Calling Leaderboard tests whether a model calls tools correctly: choosing the right function, filling arguments, handling multiple or parallel calls, and knowing when no call is needed, in single and multi-turn settings. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads BFCL Overall?
- Claude Opus 4.5 leads BFCL Overall with 77.5% as of 25 Oct 2026, ahead of Claude Sonnet 4.5 at 73.2%.
- How many models have BFCL Overall results?
- 83 model configurations have a BFCL Overall result on BenchLeader, all taken from Berkeley Function Calling Leaderboard.
- Who runs BFCL Overall and how often is it updated?
- BFCL Overall is published by Berkeley Function Calling Leaderboard. BenchLeader re-reads the published results once a day and records the date each result was published.
- Does BFCL Overall count toward the BenchLeader Index?
- Yes. BFCL Overall contributes to the agents & tools category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.