Why this benchmark is useful
Tool calling is the plumbing behind agents, automations, and research assistants. BFCL gives that plumbing a public test instead of treating every function-call demo as evidence.
Berkeley Function Calling Leaderboard
BFCL
Whether a model can choose and format function calls correctly across single-turn, multi-turn, parallel, and live API-style tool-use tasks.
What this does not measure
- Whether the downstream tool result is useful — BFCL focuses on choosing and formatting calls, not on the whole product workflow after the call returns.
- Security of tool execution — permission boundaries, prompt-injection resistance, and audit logs still need product-specific tests.
- Long-horizon agent reliability — passing function-call tasks does not prove the system can recover over dozens of steps.
Coverage map
- Task family
- Tool use
- Format
- Function/tool-calling tasks with schema-constrained expected outputs across simple, multi-turn, parallel, and live-style settings.
- Scoring
- Accuracy-style score by BFCL category and version. The exact category matters more than the aggregate.
- Maintainer
- Berkeley Gorilla
How to read the scores
Compare models within the same BFCL version and category. A model can be strong at simple calls and weak at parallel or multi-turn calls.
What it does not cover
- Whether the downstream tool result is useful — BFCL focuses on choosing and formatting calls, not on the whole product workflow after the call returns.
- Security of tool execution — permission boundaries, prompt-injection resistance, and audit logs still need product-specific tests.
- Long-horizon agent reliability — passing function-call tasks does not prove the system can recover over dozens of steps.
Scores
20 rowsEvidence ledger
2020 rows
88.7%best score
0source-checked
1sources
2026-04-15source date
Where the rows land
0255075100
Normalized to this benchmark's axis (0–100). Open the table for raw units.
Newest receipts
- Claude Sonnet 4.688.7% · Berkeley Function Calling Leaderboard V4
- GPT 5.587.4% · Berkeley Function Calling Leaderboard V4
- Claude Opus 4.886.9% · Berkeley Function Calling Leaderboard V4
- Kimi K2.685.2% · Berkeley Function Calling Leaderboard V4
- GPT 5.483.8% · Berkeley Function Calling Leaderboard V4
Scores use this benchmark's own unit and axis, not a universal quality score.
| # | Model | Release date | Score | Provenance | Trust |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 4.6AnthropicV4 aggregate score; drill into multi-turn and parallel sub-scores before choosing. | 2026-02-17 | 88.7% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 2 | GPT 5.5OpenAI | 2026-04-23 | 87.4% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 3 | Claude Opus 4.8Anthropic | 2026-05-28 | 86.9% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 4 | Kimi K2.6Moonshot | 2026-04-20 | 85.2% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 5 | GPT 5.4OpenAI | 2026-03-05 | 83.8% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 6 | Gemini 3.5 FlashGoogle | 2026-05-19 | 82.3% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 7 | Gemini 3.1 ProGoogle | 2026-02-19 | 81.6% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 8 | Mistral Large 3Mistral | 2025-12-02 | 80.1% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 9 | Gemini 3 Flash PreviewGoogle | 2025-12-17 | 79.4% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 10 | Qwen 3.7 MaxAlibaba | 2026-05-20 | 78.4% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 11 | o3OpenAIStrong in simple single-turn; drops in multi-turn complexity. | 2025-04-16 | 78.1% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 12 | Grok 4.3xAI | 2026-04-30 | 77.2% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 13 | DeepSeek V4 Flash MaxDeepSeek | 2026-04-24 | 76.8% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 14 | MiniMax M3MiniMax | 2026-05-31 | 76.2% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 15 | Llama 4 MaverickMeta | 2026-04-05 | 75.9% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 16 | DeepSeek V4DeepSeek | 2026-04-24 | 74.6% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 17 | Llama 4 ScoutMeta | 2026-04-05 | 74.1% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 18 | DeepSeek R1DeepSeek | 2025-01-20 | 72.4% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 19 | Claude Haiku 4.5Anthropic | 2025-10-01 | 71.8% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
| 20 | Cohere Command ACohere | 2026-03-15 | 70.6% | Berkeley Function Calling Leaderboard V42026-04-15 | Needs audit |
What it covers
Primary source is the Berkeley Gorilla leaderboard. VerdictPal should store category and BFCL version with every score row.
Where it breaks down
The leaderboard does not replace a product-level integration test with your schemas, auth model, error handling, and adversarial inputs.
Tools that report it
Receipts
Sources and further reading
- Berkeley Function Calling Leaderboard V4· checked 2026-06-03
- Gorilla repository· checked 2026-06-03
- BFCL leaderboard
Questions about this benchmark
Every answer below is assembled from the dated fields on this page. Nothing is written separately for search.
What does BFCL measure?
Whether a model can choose and format function calls correctly across single-turn, multi-turn, parallel, and live API-style tool-use tasks.
What does a high BFCL score not prove?
A strong BFCL result says nothing about:
- Whether the downstream tool result is useful — BFCL focuses on choosing and formatting calls, not on the whole product workflow after the call returns.
- Security of tool execution — permission boundaries, prompt-injection resistance, and audit logs still need product-specific tests.
- Long-horizon agent reliability — passing function-call tasks does not prove the system can recover over dozens of steps.
How is BFCL scored?
Accuracy-style score by BFCL category and version. The exact category matters more than the aggregate. Task format: Function/tool-calling tasks with schema-constrained expected outputs across simple, multi-turn, parallel, and live-style settings.
Is BFCL saturated?
BFCL is currently marked Active in the atlas.
Can BFCL results be contaminated by training data?
Contamination risk for BFCL is graded Medium contamination. Treat every row on this page as a public claim with a source and a date, not as a controlled experiment.
The Pack · Editorial newsletter
New tools in your inbox. Free.
One short email when a tool ships or changes status. No tracking, no third-party analytics. Unsubscribe in one click.