MCPGen - Workflow Reconstruction: leaderboard
MCPGen (Yang et al., Shanghai Jiao Tong University, 2026) builds 100 executable Model Context Protocol projects from n8n community workflow templates. Here the model receives the tool schemas and a shuffled, anonymized workflow and must recover step order, dependencies and branches; the score in percent weights step order 0.4, path robustness 0.4 and dependency recovery 0.2, averaged over five shuffles, single-turn at temperature 0, higher is better.
Metric: Reconstruction score (%; 0.4 x step order + 0.2 x dependency-edge recovery + 0.4 x path robustness, mean over five shuffled, anonymized variants of each project workflow; single-turn generation, temperature 0, 100 executable MCP projects from 16 domains). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 88.5 | #141 |
| 2 | Gemini 3 Pro (Preview) | 86.7 | #111 |
| 3 | Claude Sonnet 4.5 | 85.9 | #244 |
| 4 | Claude Haiku 4.5 (20251001) | 84 | #529 |
| 5 | GPT-5.2 | 83.7 | #172 |
| 6 | O3 | 83.7 | #224 |
| 7 | GPT-5.1 | 81.5 | #225 |
| 8 | MiniMax-M2.1 | 78.2 | #569 |
| 9 | Qwen 3 Next 80B A3B Instruct | 75.7 | #763 |
| 10 | Qwen 3 Coder 30B A3B Instruct | 75.1 | #973 |
| 11 | GPT-4o Mini | 70.7 | #1627 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mcpgen-workflow-reconstruction · How It Works · Data refreshed daily, snapshot 2026-11-12.