MCPEvol-Bench: leaderboard

Metric: Evolutionary Competency Score (per task, mean minus standard deviation of the Task Fulfillment score across the three server versions, averaged over tasks; about -1.2 to 10; 201 synthesized multi-server tasks over a fixed set of candidate MCP servers from 123 real servers; rubric scores on a 1 to 10 scale from a DeepSeek-Chat judge over the instruction, final answer and execution trajectory (MCP-Bench evaluation framework)). Source: arxiv.org. Saturation forecast: Around November 2027. 12 models tracked.

Top models

#ModelScore
1Claude Opus 4.66.09
2Claude Sonnet 4.65.22
3GPT-5.45.2
4Claude Sonnet 4.5 (Thinking)4.52
5Gemma 4 31B (IT)4.45
6GPT-5.14.32
7O34.24
8Gemini 2.5 Pro3.84
9GPT-4o3.8
10Qwen 3.5 27B3.78
11Llama 3.3 70B3.56
12Qwen 3.5 9B3.2

Interactive version: theaggregate.ai/benchmark?slug=mcpevol-bench · How It Works · Data refreshed daily, snapshot 2026-09-29.