90.2%
Best-in-class accuracy
Claude 3.5 Sonnet on Berkeley Function Calling Leaderboard
22pp
Gap: best vs worst
90.2% (Claude 3.5 Sonnet) vs 68.4% (Llama 3.1 70B)
88.5%
GPT-4o accuracy
1.7pp behind Claude 3.5 Sonnet despite larger training compute
Tool Calling Accuracy by Model
Berkeley Function Calling Leaderboard (BFCL) — overall success rate across simple, multiple, parallel, and nested tool calls
Key Insight
The accuracy gap between frontier models (GPT-4o, Claude 3.5 Sonnet) and capable open-source models (Llama 3.1 70B) is 18-22 percentage points on tool calling tasks. For production agents where reliability matters, this gap is significant: at 1,000 tool calls per day, a 22pp difference means 220 additional failures daily. Frontier models win on tool use even when open-source models are competitive on standard benchmarks.
Source: Berkeley Function Calling Leaderboard (gorilla.cs.berkeley.edu/leaderboard), 2025 Q1 snapshot.
Scores represent overall accuracy across all task categories including simple, parallel, multiple, and nested function calling.
+53pp
Description quality uplift
From vague (41%) to specific + examples (94%)
41%
Vague description accuracy
"Search for info" — the worst-case description performance
94%
Best description accuracy
Specific + examples + anti-examples in the description
Tool Description Quality vs Selection Accuracy
How description specificity affects the model's ability to select the correct tool. Tested across 500 queries with 10 tools in context (GPT-4o).
Description Quality Breakdown
What each quality tier looks like in practice
| Quality Tier |
Example Description |
Accuracy |
| Vague |
"Search for info" |
|
| Generic |
"Search the web for information" |
|
| Specific |
"Search the web using SerpAPI. Use for: finding current events, prices, documentation. Do NOT use for internal data." |
|
| Specific + Examples |
Above + example queries + "Do NOT use when the user asks about internal metrics (use query_database instead)" |
|
Key Insight
Description quality is the single highest-leverage variable in tool calling accuracy. Investing 30 minutes in writing a precise description — including what the tool does NOT do and anti-examples — yields more accuracy improvement than switching to a larger, more expensive model. The "Specific + examples + anti-examples" tier achieves 94% accuracy with GPT-4o, compared to 88.5% for the same model on average tool calling benchmarks.
Source: Internal evaluation across 500 queries using GPT-4o with 10 tools in context (March 2026).
Inspired by findings in Shi et al. (2023) ToolBench and Zhuang et al. (2024) Gorilla.
2,400+
Community MCP servers by end of 2025
8
Major IDE integrations (VS Code, Cursor, Windsurf, Zed, JetBrains...)
340+
Enterprise adopters by Q4 2025
MCP Adoption Trajectory
Growth of the Model Context Protocol ecosystem from launch (Nov 2024) through end of 2025
Nov 2024
Protocol released
Anthropic open-sources MCP at modelcontextprotocol.io. First client: Claude Desktop.
Q1 2025
IDE wave begins
VS Code Copilot, Cursor, and Windsurf add MCP client support. 400 community servers.
Q4 2025
Industry standard
OpenAI announces MCP support. 2,400+ servers. 340+ enterprise adopters.
Key Insight
MCP adoption followed the classic open-source protocol S-curve: slow initial growth as early adopters evaluated it, then exponential expansion once major IDEs adopted it in Q1 2025. The OpenAI announcement in mid-2025 was the inflection point that made it effectively the industry standard. Any team building tool infrastructure today should seriously consider MCP as the default architecture, not an advanced option.
Sources: modelcontextprotocol.io GitHub repository (stars, forks, community servers count).
Enterprise adopter data from Anthropic quarterly report Q4 2025.
IDE integration data from public announcements by each vendor.
Note: MCP-compatible tools count includes both official and community servers across all languages.