77.5%
#1
overall BFCL accuracy of Claude-Opus-4.5 in function-calling mode, early 2026; Claude-Sonnet-4.5 73.2%

What the chart shows

Function calling lets a model use external tools by writing structured requests for another system to run. The leaderboard’s fourth version gives agentic tasks such as web search and memory 40% of the score and multi-turn dialogues 30%. In early 2026 Claude-Opus-4.5 led with 77.5%, ahead of Claude-Sonnet-4.5 (73.2%) and Gemini-3-Pro-Preview in prompt mode (72.5%).

The top 15 entries span 21.2 points, down to 56.2% for Gemini-2.5-Flash. The same model can score differently by mode: Grok-4-0709 scored 63.0% with prompting and 61.4% with native function calling.

Key numbers

MeasureValue
Claude-Opus-4-5-20251101 (FC)77.47%
Claude-Sonnet-4-5-20250929 (FC)73.24%
Gemini-3-Pro-Preview (Prompt)72.51%
Gemini-3-Pro-Preview (FC)68.14%
Gemini-2.5-Flash (FC), 15th56.24%

How to read it

  • FC means the model’s native function-calling interface; Prompt means the tools are described in the prompt. Some models appear twice, once per mode.
  • Snapshot of the top 15 entries on the leaderboard (gorilla.cs.berkeley.edu) in early 2026.
  • Overall accuracy is a weighted average across agentic, multi-turn, live, non-live and hallucination categories.
Source: Berkeley Function Calling Leaderboard, 2026, via the AI Index Report 2026 (Stanford HAI), figure 2.2.3: Berkeley Function Calling: overall accuracy. Chart and text by AI Adoption.
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart